Computational Field Theory and Pattern Formation 1. Monte Carlo Methods 1.1. Introduction

Transcription

Computational Field Theory and Pattern Formation 1. Monte Carlo Methods 1.1. Introduction
Computational Field Theory
and Pattern Formation
Raul Toral
Departament de Fsica, Universitat de les Illes Balears, 07071-Palma de
Mallorca, Spain
1. Monte Carlo Methods
1.1. Introduction
In many occasions we need to compute the value of a high-dimensional integral
of the form:
Z 1
1
dx1 : : :
dx f (x : : : xN )G(x1 : : : xN )
;1
;1 N x^ 1
Z
Z
dXfx^ (X)G(X) (1:1)
here N is a large number and we have denoted by X the set (x1 : : : xN ) of real
variables. The function fx^ (X) is non-negative and normalized, i.e.:
(i)
fx^ (X) 0
Z
Z 1
Z 1
(1:2)
(ii)
fx^ (X) dX
dx1 : : :
dxN fx^ (x1 : : : xN ) = 1
;1
;1
In next chapters we will give specic examples of when one needs to perform
this type of calculation. If the above integral can not be computed by analytical methods (which happens more often than desired), one has to resource to
numerical methods. One must realize, however, that a simple extension of the
standard methods used in small{N numerical integration (Simpson rules, Gauss
integration, etc. (Faires et al. 1993)) to compute numerically the value of the
integral will fail for very high{dimensional integrals. For instance, let us take
N = 100 (and this is a rather small number for typical applications) and suppose we want to use Simpson methods. We need to generate a grid in the X space
and sum up all the values of G(X) on the points of this grid. If we choose, say,
a grid of 10 points per every coordinate, we will have 10100 terms to add in the
Simpson rule. If every addition takes 10;12 s (and today's computers take much
longer than that) the time needed for the calculation exceeds any astronomical
unit of time (including the age of the universe). The situation does not improve
2
Raul Toral
by diminishing the number of points in the integration grid. Even 2 points per
integration variable is beyond the capability of any present (or future) computer.
The solution to this problem is given by application of the Monte Carlo
techniques (Kalos et al. 1986 Binder 1992, Heermann 1986). For our particular
example, they consist in interpreting fx^ (X) as a probability density function
of some random variable ^x and, consequently, the integral is nothing but the
average value of the function G(X),
hGi =
Z
dXfx^ (X)G(X)
(1:3)
The key idea behind Monte Carlo integration is to approximate the previous
average by a sample average:
hGi M G] = M1
M
X
k=1
G(X (k) )
(1:4)
where X (k) k = 1 : : : M are values of the N{dimensional variable X distributed according to the probability density function fx^ (X). This is called the
technique of importance sampling. The basic principle of sampling is the same
one that allows politicians to have an accurate knowledge of the election results
before election day by making random polls amongst a representative sample of
the electors.
A more precise statement on the relation between the average hGi and the
sample average M G] is given by the central limit theorem (Feller 1971). This
theorem states that, if the values of X (k) are statistically independent, then the
sample average M tends in the limit M ! 1 to a Gaussian distribution of
mean hGi and variance 2 M ] given by:
2M ] = M1 2 G] = M1 hG2 i ; hGi2
(1:5)
It is costumery to express the relation between M G] and hGi as:
hGi = M G] M ]
(1:6)
which is to be interpreted in the usual statistical sense for the Gaussian
distribution, i.e. that there is a 95:45% probability that hGi lies in the interval (M G] ; 2M ] M G] + 2M ]), 99:73% in the interval (M G] ;
3M ] M G] + 3M ]), etc. In practice, the unknown variance 2 G] can be
replaced by the sample variance:
!2
M
M
1 X
(k) )2 ; 1 X G(X (k) )
G(X
(1:7)
M k=1
M k=1
According to relation (1.5) the error in a Monte Carlo integration decreases as
the inverse square root of the number of samplings M. In the case that the values
X (k) are not independent, two major modications are needed:
2 G] Computational Field Theory and Pattern Formation
3
(a) The variance of the sample average is given by:
2
2 M ] = MG] (2G + 1)
(1:8)
where the autocorrelation time G is given (in the limit of large M) by:
G =
1
X
k=1
G (k)
(1:9)
here the normalized autocorrelation function G (k) is dened as:
(k )
(k+k ) )i ; hGi2
G (k) = hG(X )G(X
2 G]
0
0
(1:10)
The situation for non zero autocorrelation time can arise, for instance,
when the value X (k) depends on the value X (k;1) . In our poll example,
this can occur if the k{th person to be questioned lives in the neighbourhood of the (k ; 1){th person, such that their social classes are likely to
be similar. Intuitively, G measures the number of values of the sample
X (k) X (k+1) : : : X (k+ G ) that we have to discard in order to consider that
X (k) and X (k+ G ) are independent of each other. A problem of Monte Carlo
methods is that, in many cases of interest, G becomes very large and grows
with the number of variables N.
(b) The relation hGi = M G] M ] has to be interpreted now according to
Chebichev's theorem (Feller 1971):
P(jM G] ; hGij > kM ]) > 1 ; k12
(1:11)
i.e. the probability that hGi lies in the interval (M G] ; 2M ] M G] +
2M ]) is at least 75%, that it lies in (M G] ; 3M ] M G] + 3M ]),
at least 88:89%, etc.
The above Monte Carlo importance sampling procedure is one of the most
powerful methods available to compute high{dimensional integrals. In order to
completely specify the method, though, we need a procedure to generate the
values of the samples X (k) that appear in the above formulae. These should be
values of the N{dimensional variable (x1 : : : xN ) distributed according to the
probability density function fx^ (X). There are several methods devised to generate the required values X (k) . Before we can explain the very powerful methods
available we need to study in detail some simpler cases.
4
Raul Toral
1.2. Uniform sampling
Let us start with a simple example. Consider the one{dimensional (N=1) integral:
I=
Z
0
1
dx cos(x)
(1:12)
this is a particular case of Eq.(1.1) in which the function fx^ (x) is equal to 1 in the
interval (0 1) and 0 otherwise. The name uniform sampling comes from the
fact that the variable x is equally likely to take any value in the interval (0 1) (or,
generally speaking, in any nite interval (a b)). Although one could, in principle,
devise some physical process that would produce values of a variable x uniformly
distributed in the interval (0 1), it is preferable, for many reasons, to use simple
algorithms that can be programed on a computer and that produce "random{
type" (or pseudo{random) numbers uniformly distributed in the interval (0 1).
We will not discuss in these lectures the interesting topic of pseudo{random
number generation. An excellent exposition can be found in references (Knuth
1981) and (James 1990). Let us simply assume that there exists a fortran function, which we will call ran u(), which returns an independent random number
uniformly distributed in the interval (0 1). By using such a function, we can
implement very easily the algorithm of uniform sampling:
99
m=1000
r=0.0
s=0.0
do 99 k=1,m
x=ran_u()
gk=g(x)
r=r+gk
s=s+gk*gk
continue
r=r/m
s=s/m-r*r
s=sqrt(s/m)
write(6,*) r,'+/-',s
end
function g(x)
g=cos(x)
return
end
The reader should nd out the name given to the function ran u() is his
own computer and run the above program checking that the error in the output
value scales as the inverse square root of the number of samples M.
Computational Field Theory and Pattern Formation
5
1.3. Importance sampling for N=1
In the N=1 case, when the variable X = x has only one component, a simple and
important theorem allows to generate x(k) very easily. Recall that the probability
distribution function Fx^(x) is dened by integration of the probability density
function fx^ (x) as:
Z x
Fx^(x) =
fx^ (y) dy
(1:13)
;1
Theorem 1 If fx^ (x) is a one-variable probability distribution function, then the
variable u = Fx^(x) is uniformly distributed in the interval (0 1).
As a consequence, we have the
Corollary 1 If u is uniformly distributed in the interval (0 1), then x = Fx^;1(u)
is distributed according to fx^ (x).
(The proof of this theorem is left as an exercise for the reader). This theorem reduces the problem of generating random numbers according to a given
distribution fx^ (x) to the inversion of the corresponding probability distribution
function. If, for instance, we want to use the method of importance sampling to
compute the integral
Z 1
I=
dxe;xx2
(1:14)
0
we can take fx^ (x) = e;x if x 0, and G(x) = x2. This choice for fx^ (x) respects
positivity and normalization. To generate values of x(k) according to fx^ (x) we
need to invert Fx^ (x), which in this case can be done analytically:
Fx^(x) =
x
Z
0
dye;y = 1 ; e;x = u ! x = Fx^;1(u) = ; log(1 ; u)
(1:15)
Since 1 ; u is also uniformly distributed in the interval (0 1) we can write simply:
x = ; log(u)
(1:16)
which is equivalent to x = ; log(1 ; u) from the statistical point of view. A
program to compute integral (1.14) can be:
99
m=1000
r=0.0
s=0.0
do 99 k=1,m
x=ran_f()
gk=g(x)
r=r+gk
s=s+gk*gk
continue
6
Raul Toral
r=r/m
s=s/m-r*r
s=sqrt(s/m)
write(6,*) r,'+/-',s
end
function g(x)
g=x*x
return
end
function ran_f()
ran_f=-log(ran_u())
return
end
A technical point is that, in some cases, the function Fx^;1(u) is not expressible
in terms of elementary functions and some kind of approximative methods (such
as numerical interpolation) might be needed to compute it. It is important to
realize that the decomposition of a given integral in fx^ (x) and G(x) to put it in
the form of Eq.(1.3) might not be unique. In the previous example, Eq.(1.14),
we could have taken as well fx^ (x) = xe;x and G(x) = x. This choice, however,
makes the probability distribution function Fx^ (x) dicult to invert.
An important case in which the function Fx^;1(u) is not easily calculable is
that of Gaussian random numbers (of mean zero and variance 1) for which the
probability density function is:
2
(1:17)
fx^ (x) = p1 e;x =2
2
(random numbers y of mean and variance 2 are easily obtained by the linear
transformation y = x + ). The inverse probability distribution function is
p
x = 2 erf ;1 (2u ; 1)
(1:18)
x1 =p cos()
x2 =p sin()
(1:19)
where erf ;1 (z) is the inverse error function (Abramowitz et al. 1972 Valls
et al. 1986). The inverse error function does not usually belong to the set of
predened functions in a programming language, although some libraries (for
example, the NAG library) do include it in their list of functions. An alternative
to the generation of Gaussian distributed random numbers is the algorithm of
Box-Muller-Wiener (Box et al. 1958 Ahrens et al. 1972 1988) which is based
on a change of variables to polar coordinates. Namely,
If x1 and x2 are independent Gaussian variables, it is easy to prove that and are independent variables. follows an exponential distribution and a uniform
Computational Field Theory and Pattern Formation
7
distribution, so that their inverse distribution functions can be expressed in
terms of elementary functions. Namely,
f () = 12 e;=2 ! F () = 1 ; e;=2 0
1 ! F () = 0 2
(1:20)
f () = 2
2
The resulting algorithm is:
p
x1 = ;2 log(u) cos(2v)
(1:21)
p
x2 = ;2 log(u) sin(2v)
Where u and v are independent random variables uniformly distributed in the
interval (0 1). The main advantage of this Box-Muller-Wiener algorithm is that
it is exact, yielding two independent Gaussian variables, x1 x2 from two independent uniform variables, u v. Its main disadvantage, though, is that it is
extremely slow since it involves the calculation of trigonometric, logarithm and
square root functions. In most of the applications a linear interpolation approximation to the inverse error function does produce suciently good quality
Gaussian random numbers at a considerable gain in speed (Toral et al. 1993). A
possible implementation of the Box-Muller-Wiener algorithm is the following:
function ran_g()
data is /-1/
is=-is
if (is.eq.1) then
a=sqrt(-2.0*log(ran_u()))
b=6.2831853072*ran_u()
ran_g=a*cos(b)
x2=a*sin(b)
return
endif
ran_g=x2
return
end
(a practical note: it might be necessary to tell the compiler that the two functions
ran u() that appear in the previous program produce dierent values and need
to be computed separately, sometimes compilers are too clever).
Another important case that deserves being considered explicitly, in despite of
its simplicity, concerns the generation of events with a given probability. Imagine
we want to simulate tossing a biased coin, with a p = 0:6 probability for heads
(and 0:4 for tails). We need to generate a variable that takes some value (1, for
example) 60% of the times and another value (0), 40% of the times. This can
be done by comparing a random number u uniformly distributed in the interval
(0 1) with the given probability p. If u p then we take x = 1, otherwise we
take x = 0. This is achieved by the program:
8
Raul Toral
if (ran_u().lt.p) then
x=1
else
x=0
endif
For N-dimensional variables (x1 : : : xN ), the situation is much more complex. The equivalent of Theorem 1 states that in order to generate values of the
N{dimensional variable (x1 : : : xN ) we need (Rubinstein 1981): (i) generate Nindependent random numbers (u1 : : : uN ) uniformly distributed in the interval
(0 1) (this is the easy part) and (ii) solve the following set of equations:
Fx1 (x1 ) =u1
Fx2 (x2 jx1) =u2
(1:22)
: : :: : :
FxN (xN jx1 : : : xN ;1 ) =uN
Where, for example, Fx2 (x2 jx1) is the conditional probability distribution function of the variable x2 given that x1 has taken a particular value, and so on.
The calculation of the conditional probability distribution functions is generally
at least as complicated as the calculation of the original integral we wanted to
compute numerically and the above procedure is of little practical use. In order to develop alternative methods suitable for the generation of N-dimensional
variables we need rst to introduce the so{called rejection methods for 1 variable.
1.4. Rejection method for N=1
In those cases that the inverse probability distribution function is dicult to
compute, the rejection method oers a very convenient alternative. Also, it is
the basis of the N-dimensional methods which we will develop later. The method
is based on the fact that the probability density function fx^ (x) is proportional
to the probability that the variable x takes a particular value. If, for example,
fx^ (x1 ) = 2fx^ (x2 ), we can arm that the value x1 is twice as probable as the
value x2. The rejection method (in its simplest version) proposes the values x1
and x2 with the same probability and then accepts the proposed value x with
a probability h(x) proportional to fx^ (x), such that, in our example, x1 will be
accepted twice as many times as x2. Consequently, x1 will appear twice as many
times as x2, which is equivalent to saying that x1 is twice as probable as x2
as desired. We will illustrate the method by an example. Let us consider the
probability density function:
fx^ (x) = 6x(1 ; x)
x 2 (0 1)
(1:23)
Which has the shape indicated in Fig.1.
We propose a value of x uniformly in the interval (0 1). The proposed value
has to be accepted with a probability h(x) proportional to fx^ (x), h(x) = fx^ (x).
Computational Field Theory and Pattern Formation
9
Fig. 1. Probability density function f^x (x) = 6x(1 ; x) (solid line) and acceptance
probability h(x) = 4x(1 ; x) (dashed line)
The constant is arbitrary but the resulting function h(x) has to be interpreted
as a probability, which means that we have to keep the bounds 0 h(x) 1. This
yields maxx1f^x (x) . Obviously, the rejection method is more ecient when the
acceptance probability is high, which means that one has to take the maximum
possible value for . In our example, the maximum for fx^ (x) is at x = 0:5,
fx^ (0:5) = 1:5, so that we take = 2=3 and, consequently, h(x) = 4x(1 ; x), see
Fig.1. The acceptance process is done as explained in the previous section by
comparing the probability h(x) with a random number uniformly distributed in
(0 1). The whole process is repeated until the proposed value is accepted. The
nal algorithm can be coded as:
1
function ran_f()
h(x)=4.0*x*(1.0-x)
x=ran_u()
if (ran_u().gt.h(x)) goto 1
ran_f=x
return
end
We will now consider a more general version of the rejection method in which
a value is proposed according to some probability distribution function gx^ (x) and
then accepted according to some probability h(x). The probability distribution
function of the joint process "propose the value x and accept it" is gx^(x)h(x).
According to the Bayes theorem, the probability density function of the variable
x given that it has been accepted is (Grimmett et al. 1982):
fx^ (x) = R 1 gx^ (x)h(x)
(1:24)
;1 dzgx^(z)h(z)
10
Raul Toral
One interesting thing to notice is that the previous expression is properly normalized. This is very important in those cases where the normalization constant
for fx^ (x) is not known explicitly. Looking at Eq.(1.24) one can see that there is
some freedom in choosing the functions gx^(x) and h(x). The only requirement
is that its product gx^ (x)h(x) be proportional to the function fx^ (x) we want to
generate. An important constraint is that the proposal gx^ (x) must be normalized and that the acceptance function h(x) must be bound in the interval (0 1).
For instance, in the example of Eq.(1.23), we could have considered gx^ (x) = 2x.
The function h(x) should then be proportional to 3(1 ; x). Since 0 h(x) 1
one takes simply h(x) = 1 ; x. The generation of gx^(x) = 2x is straightforward
since the inverse probability distribution function is G;x^ 1 (u) = pu. The proposed value is accepted with probability 1 ; x, i.e. rejected with probability x.
This can be coded as:
1
function ran_f()
x=sqrt(ran_u())
if(ran_u().lt.x) goto 1
ran_f=x
return
end
An ecient method is one that requires, on average, a small average number
of proposals before acceptance. This average number is given by the inverse of
the overall acceptance probability, pa which is nothing but the denominator of
expression (1.24):
Z 1
pa p(accept) =
dzg (z)h(z)
(1:25)
;1 x^
It is easy to show that of the two methods developed for the function (1.23) the
rst algorithm is more ecient than the second.
Let us consider yet another example: to generate numbers distributed according to
2
(1:26)
fx^ (x) = C exp ; x2 ; x4
the obvious choice is (notice that the precise value of C is not needed):
2
gx^ (x) = p1 e; x2
2
(1:27)
4
;
x
h(x) =e
Here gx^ (x) is nothing but a Gaussian distribution, and h(x) is the optimal choice,
given the choice for gx^ (x), since maxx h(x) = 1. The overall acceptance probability is, according to Eq.(1.25):
Z 1
2
pa =
dz p1 e; z2 e;z4 0:62028
(1:28)
;1 2
And the average number of trials needed to accept a value is 1=0:62028 1:6.
Computational Field Theory and Pattern Formation
11
1.5. Rejection with repetition
There is a modication of the rejection method consisting in the following: if
the proposed value is not accepted then, instead of proposing a new one, the
previous value is repeated. This method obviously produces correlated values
(some of them will be equal to the previous ones) but, however, can be very
ecient for vector computers in which the usual structure of the rejection method
prevents vectorization. We still have to prove that the values generated with this
method do follow the required distribution fx^ (x). According to Bayes theorem,
the probability density function fx^n (x) generated at the n{proposal step will
have two dierent contributions corresponding to the acceptance or rejection,
respectively, of the proposed value:
fx^n (x) = fx^n (xjaccept)p(accept) + fx^n (xjreject)p(reject)
(1:29)
The rst term of the right hand side can be written as the probability of acceptance given x (which is the function h(x)) times the probability density function
of proposing x (which is gx^ (x)):
fx^n (x) = h(x)gx^ (x) + fx^n (xjreject)(1 ; pa )
(1:30)
If x has been rejected, the previous value is repeated which means that the
probability density function at the n{step in the case of rejection is the same
that the one at the (n ; 1){step:
1
fx^n (x) = h(x)gx^ (x) + fx^n 1 (x) 1 ;
dzh(z)gx^ (z)
;1
;
Z
(1:31)
where we have substituted pa as given by Eq.(1.25). The solution of this linear
recurrence relation is:
#
"
h(x)g
(x)
x
^
n
+ R 1 h(x)gx^ (x)
(1:32)
fx^n (x) = (1 ; pa ) fx^0 (x) ; R 1
;1 dzh(z)gx^ (z)
;1 dzh(z)gx^(z)
We can write this result in terms of the desired distribution fx^ (x), Eq. (1.24):
fx^n (x) = (1 ; pa )n fx^0 (x) ; fx^ (x)] + fx^ (x)
(1:33)
Given that 0 < pa 1, one concludes that the solution tends to the desired
distribution in the limit n ! 1 independently of the initial distribution fx^0 :
limn!1 fx^n (x) = fx^ (x)
(1:34)
If the initial proposed numbers x0 are distributed according to fx^ (x), i.e. if
fx^0 (x) = fx^ (x), then we would have fx^n (x) = fx^ (x) 8n and the numbers we
obtain are distributed according to the desired distribution from the beginning.
However, if the initial numbers are not distributed according to fx^ (x) (as it is
usually the case) we can still arm that, thanks to the factor (1 ; p)n , the initial
condition will be lost after a number suciently large of proposals. This process
of rejecting the, say, M0 initial values produced because they are not distributed
12
Raul Toral
yet according to the required distribution is called thermalization. The initial
time needed for thermalization can be computed from the non-linear correlation
function (Binder et al, 1988):
hx(k)i ; hxi
NL
x (k) = hx(0)i ; hxi
(1:35)
The time M0 can be measured as a characteristic time for the decay of the
above function towards the stationary value 0. For the method of rejection with
repetition the non-linear correlation function coincides with the auto-correlation
function (see later):
k
NL
(1:36)
x (k) = (1 ; pa )
The thermalization time M0 can be dened as the time it takes the above function to reach a small value .
M0 = ln(1ln; p )
(1:37)
a
If, for instance, = 10;8, pa = 0:5, we have M0 27.
We can illustrate now the enhancement of the errors due to the correlations.
In this method of rejection with repetition, one can compute the correlation
function of the variable x as (see Eq.(1.10)):
hx(k)x(k+k ) i ; hxi2 = (1 ; p )k
x (k) = klim
a
!1 hx2 i ; hxi2
From Eq.(1.9) we deduce that the autocorrelation time is:
0
0
x =
1
X
k=1
x (k) = 1 ;p p
(1:38)
(1:39)
Remember that, according to Eq.(1.8) the statistical error gets enhanced by the
factor (2x + 1)1=2.
We nally mention that it is possible to interpret this algorithm of rejection
with repetition as a Markov succession (Grimmett et al. 1982). Indeed, Eq.(1.31)
can be cast in the form:
fx^n (x) =
Z
1
f(xjy)fx^n 1 (y) dy
;1
;
(1:40)
where the transition probability is:
1
f(xjy) = h(x)gx^ (x) + 1 ;
dzh(z)gx^ (z) (x ; y)
;1
Z
(1:41)
Computational Field Theory and Pattern Formation
13
1.6. The algorithm of Metropolis et al.
Although it is very powerful, the rejection method can fail if the distribution is
very peaked around an (unknown) value x0 (see Fig.2). If our proposal does not
take into account that only a very small part of the variable space (that around
x0) is important, we are bound to have a extremely high rejection probability
yielding the rejection method completely useless. The solution comes by making
a proposal that does not dier much of the present value of the random variable we are sampling, i.e. by having the transition probabilities of the Markov
succession depend explicitly on the present value, say y, of the variable. This
implies that the functions h(x) and gx^ (x) depend on y and become two-variable
functions, gx^(xjy), h(xjy). The transition probability Eq.(1.41) then becomes
also a two variable function:
f(xjy) = h(xjy)gx^ (xjy) + 1 ;
1
h(z jy)gx^ (z jy) dz (x ; y)
;1
Z
(1:42)
Fig. 2. Function f^x(x) for which the rejection method can fail
The equation giving the evolution of the probability density function at the nstep is still Eq. (1.40), but now the recursion equation for fx^n (x) does not tell
us in a straight way which is the stationary (n ! 1) distribution. We would
like to have fx^ (x) as stationary solution of the recursion equation, i.e:
1
f(xjy)fx^ (y) dy
(1:43)
;1
R1
By using the normalization condition ;1
f(yjx) dy = 1 and the denition (1.42)
for the transition probability f(xjy), one gets easily:
fx^ (x) =
Z
Z
dy gx^ (xjy)h(xjy)fx^ (y) ; gx^(yjx)h(yjx)fx^ (x)] = 0
(1:44)
14
Raul Toral
A sucient condition to fulll this relation is given by the detailed balance
condition:
gx^ (xjy)h(xjy)fx^ (y) = gx^ (yjx)h(yjx)fx^ (x)
(1:45)
to be satised by the proposal gx^ (xjy) and the acceptance probability h(xjy). Let
us remark that if the detailed balance condition is satised and if fx^0 (x) = fx^ (x)
then fx^n (x) = fx^ (x) 8n. Also, if the transition probabilities satisfy the condition
of ergodicity, which, roughly speaking, means that any value for the random
variable can be achieved starting from any other value after a sucient number
of steps, then we can assure that fx^ (x) is the only stationary solution of the
Markov succession and that limn!1 fx^n (x) = fx^ (x), independently of the value
of fx^0 (x) (see Kalos et al, 1986).
It is important to stress that any functions gx^ (xjy) and h(xjy) that satisfy
ergodicity and the detailed balance condition are suitable to use in this rejection
method. The most widely used solution is the one given by the the Metropolis
algorithm (Metropolis et al. 1953) in which gx^(xjy) is a given function usually
of the form:
1
gx^ (xjy) = 2
if jx ; yj (1:46)
i.e. x is sample uniformly from the interval (y ; y + ). Once the value of
x has been proposed one looks for acceptance probabilities h(xjy) verifying the
detailed balance condition:
h(xjy) = gx^(yjx)fx^ (x) q(xjy)
(1:47)
h(yjx) gx^(xjy)fx^ (y)
When searching for solutions of these equations one has to remember that h(xjy)
is a probability and must also satisfy the condition 0 h(xjy) 1. A possible
solution is the Metropolis solution:
h(xjy) = min(1 q(xjy))
(1:48)
Another widely used solution is the Glauber solution (Glauber 1963):
jy)
h(xjy) = 1 +q(xq(x
(1:49)
jy)
A family of solutions is obtained by:
p
h(xjy) = q(xjy)! (q(xjy))
(1:50)
where !(z) is any function satisfying
!(z) = !(z ;1 )
(1:51)
In particular, the Metropolis solution is recovered taking:
!(z) = min(z z ;1 )
(1:52)
and the Glauber solution taking
Computational Field Theory and Pattern Formation
!(z) = z 1=2 +1 z ;1=2
15
(1:53)
Before we use the Metropolis algorithm to study some models of eld theory,
we want to nish this section by showing the use of the Metropolis algorithm
with a simple example in a one variable case. We will consider the Gaussian
distribution:
2
fx^ (x) = Ae;x =2
(1:54)
the normalization constant A is irrelevant for the Metropolis method. The proposal will be a value belonging to the interval (y ; y + ). Notice that this
choice (Eq.(1.46)) satises the symmetry relation gx^ (xjy) = gx^ (yjx). This proposal will be accepted with a probability given, for example, by the Metropolis
solution Eq.(1.48):
h(xjy) = min(1 q(xjy)) = min(1 e(y2 ;x2 )=2)
(1:55)
i.e 2if j2xj jyj accept with probability 1, otherwise accept with probability
e(y ;x )=2 . The Metropolis algorithm can be coded as:
function ran_g(y,delta)
x=y+delta*(2*ran_u()-1)
if(abs(x).gt.abs(y)) then
if (exp(0.5*(y-x)*(y+x)).lt.ran_u()) then
ran_g=y
return
endif
endif
y=x
ran_g=y
return
end
Intuitively, the eect of this algorithm is to accept with a larger probability those proposals which tend towards the origin, where the probability has a
maximum.
An important warning concerns the fact that we must keep the old value of
the variable if the new one has been rejected. We can not keep proposing until
a value is accepted (at variance with what happened in the ordinary rejection
method in which the two procedures were correct). If one does not keep the old
value, a reasoning similar to the one used in the rejection method, section 1.4,
leads to the following transition probabilities, see Eq.(1.24):
f(xjy) = R 1 h(xjy)g(xjy)
(1:56)
;1 dzh(z jy)g(z jy)
and:
f(yjx) = R 1 h(yjx)g(yjx)
(1:57)
;1 dzh(z jx)g(z jx)
16
Raul Toral
Now it is very dicult to satisfy the detailed balance conditions, Eq.(1.45),
namely:
h(xjy)g(xjy) f (y) = R h(yjx)g(yjx) f (x)
R1
(1:58)
x
^
x^
1
;1 dzh(z jy)g(z jy)
;1 dzh(z jx)g(z jx)
since the integrals are generally dicult to compute (in particular, the Metropolis solution does not satisfy this condition). The reader can test the following
modication of the Metropolis algorithm for the Gaussian distribution:
1
function ran_g(y,delta)
x=y+delta*(2*ran_u()-1)
if(abs(x).gt.abs(y)) then
if (exp(0.5*(y-x)*(y+x)).lt.ran_u()) goto 1
endif
y=x
ran_g=y
return
end
and check that it does not produce Gaussian random numbers (it is worse for
large ).
For eciency of the Metropolis algorithm must be chosen such that the
autocorrelation time of the algorithm, , takes its minimum value. It is easy to
understand why there must be a minimum of as a function of . If is very
small, the new proposal will be very close to the old one (hence highly correlated),
will be large and h(xjy) will be very close to 1 such that the acceptance will be
also close to 1. On the other hand, if is large, the acceptance probability will
be small and the proposed value will be often rejected, such that the new value
will be equal to the old one (again highly correlated). A rule of thumb tells us
that must be chosen such that the acceptance probability is neither too high,
neither too small, i.e. of the order of 50%.
1.7. Rejection method in the N{dimensional case
We really do not need to develop any new concepts. The nice thing about the
Metropolis algorithm is that is easily generalizable to the N-dimensional case.
Simply replace x and y in the above expressions by the N-dimensional variable
vectors (x1 : : : xN ) and (y1 : : : yN ), respectively, and all of the above formulae
and concepts will still apply. In order to keep the acceptance probability within
reasonable limits, the proposal can not be too dierent from the old value. This
is achieved by changing only one of the N variables at one time. The variable
to be updated can be chosen randomly amongst the set of the N variables or
sequentially. If chosen randomly the function gx^(xjy) is explicitly given by:
N 1 Y
X
gx^ (xjy) = N1
2 (xj ; yj )
i=1
j 6=i
(1:59)
Computational Field Theory and Pattern Formation
17
Since it veries the symmetry condition gx^(xjy) = gx^ (yjx), the function
q(xjy) is simply q(xjy) = fx^ (x)=fx^ (y). The acceptance probability can be chosen
as the Metropolis solution h(xjy) = min(1 q(xjy)) or other solution.
1.8. Heat Bath
In the Metropolis algorithm the acceptance probability h(xjy) was determined
(with some freedom) once the proposal gx^ (xjy) had been specied. In the so
called heat bath algorithm, which is useful for N-dimensional variables, the
proposed value for the variable is always accepted, i.e. one chooses
h(xjy) = 1
(1:60)
The detailed balance condition becomes
gx^ (xjy)fx^ (y) = gx^ (yjx)fx^ (x)
(1:61)
In the 1-variable case, a trivial solution (remember that gx^ (xjy) must be a probability density function) is: gx^ (xjy) = fx^ (x) which is independent of the old
value y. In the N-variable case, a solution is found by changing only one of the
N variables at a time:
N
X
Y
gx^ (xjy) = N1 gx^(xi jy) (xj ; yj )
(1:62)
i=1
j 6=i
where each of the functions gx^ (xi jy) satises detailed balance:
gx^ (xi jy)fx^ (x1 : : : xi;1 yi xi+1 : : : xN ) =
(1:63)
gx^ (yi jx)fx^(x1 : : : xi;1 xi xi+1 : : : xN )
A solution of this functional equation is obtained by taking gx^ (xi jy) = gx^ (xi),
independent of y, as the following conditional probability:
gx^ (xi ) = fx^ (xi jx1 : : : xi;1 xi+1 : : : xN )
(1:64)
It is trivial to verify detailed balance, Eq.(1.63), if one remembers the denition
of the conditional probability:
fx^ (xi jx1 : : : xi;1 xi+1 : : : xN ) = fx^f(x(x1 ::::: :xix;1 x ix xi+1 ::::: :x xN) ) (1:65)
x
^ 1
i;1 i+1
N
Intuitively, what the heat bath method does when updating variable xi, is
to select the value of xi according to the conditional probability given that
the rest of the variables is xed. These xed variables act a a heat bath
for variable xi. In many cases, it turns out that the conditional probability
gx^ (xi) = fx^ (xi jx1 : : : xi;1 xi+1 : : : xN ) takes a simple form and can be sampled by any of the 1-variable methods explained in this chapter.
18
Raul Toral
2.
4
Model
2.1. Introduction and basic denitions
We will illustrate the use of Monte Carlo techniques in eld theory with the
scalar 4 model. This model has been used in many dierent contexts, in such
dierent topics as quantum eld theory (Zinn Justin 1989) or in the study of
structural phase transitions (Cowley 1980 Bruce 1980 Bruce et al. 1980).
Let us consider a d-dimensional (hyper-cubic) regular lattice , consisting
of N = Ld points. Every point i = 1 : : : N of this lattice has d coordinates:
i = (j1 : : : jd ). Periodic boundary conditions are assumed on this lattice. On
every site of the lattice there is a scalar variable i . The set of all variables is
] (1 : : : N ). We also introduce a Hamiltonian function H given by:
H(]) =
N
X
ad
"
d ; 2
;b 2 + u 4 + K X
i
i
#
(2:1)
2 =1
a0
The sum over runs over d-nearest neighbours of site i, i.e. if the coordinates
of i are (j1 : : : jd ) then the coordinates of i are (j1 : : : j + 1 : : : jd ). The
continuum limit of the system can be obtained by letting the lattice spacing a0
tend to 0 and the system size L to 1. In this limit, the sums are replaced by
integrals and the sum over tends to a gradient. The continuum Hamiltonian
is then:
Z
u
K
;
b
2
4
2
(2:2)
H(]) = dr 2 (r) + 4 (r) + 2 jr(r)j
i=1
0
2 i
4 i
This expression for the Hamiltonian is the preferred one for analytical treatment,
such as series expansions or renormalization{group treatments in momentum
space (Amit 1984). The lattice version is the preferred one for numerical studies,
besides the fact that it yields a regularization of the Hamiltonian (Parisi 1988).
We will consider from now on only the lattice version of the 4 model. We can
set the lattice spacing a0 = 1, since it can be rescaled in the parameters b u and
K.
The rst two terms of the sum appearing in the Hamiltonian (2.1) are local
terms (depending only on the eld at location i) and can be thought of as local
potential terms V (i ):
V (i ) = ;2b 2i + u4 4i
(2:3)
For stability of this local potential, the parameter u must be greater than 0.
The local potential changes qualitative when the parameter b changes sign (see
Fig.3). If b < 0 the local potential has only one minimum at i = 0. On thepother
hand, when b > 0 there are two minima of equal depth located at i = b=u.
The third term in (2.1), the one multiplied by K, is called the interaction
term. When K > 0, the contribution of this term is always non negative and
the ground state of the Hamiltonian H is reached when the contribution of this
term is zero, i.e. when all the elds take the same values i = 0 8i. This is
Computational Field Theory and Pattern Formation
19
the situation considered by the simple mean eld theory (Landau et al. 1980).
Since all the elds take the same value, the mean eld Hamiltonian reduces to:
N ;1 HMF (0) = ;2b 20 + u4 40
(2:4)
Fig. 3. Local Potential V ( ), Eq.(2.3), in the cases (i) b > 0 and (ii) b < 0
The value of the mean eld value 0 depends on the constants b u and is
given by:
8
>
0
if
b<0
<
0 = > r b
(2:5)
:
if
b
>
0
u
(see Fig.4). To go beyond mean eld theory one needs to consider uctuations:
the elds i do not take a constant value, but uctuate from one lattice site
to another. In this case, we need a statistical description which, according to
the principles of Statistical Mechanics (Pathria 1972), is given by the canonical
distribution. The probability density function f(]) governing the statistical
properties of the elds ] at inverse temperature is:
;H]
f(]) = Ze(b u K)
(2:6)
The denominator of this expression is called the partition function and it is
the multiple integral of the numerator for all the eld values:
Z (b u K) =
Z 1
1
d1 : : :
d e;H(])
;1
;1 N
Z
Z
de;H(])
(2:7)
The magnitudes of interest are computed as averages of eld functions, G(]),
with the probability density function f(]):
20
Raul Toral
Z
hG(])i = Z ;1 dG(])e;H(])
(2:8)
Example of quantities of interest are the magnetization, m dened as:
N
X
m = h N1 i i
(2:9)
i=1
the energy, e:
H
]
e= N
the magnetic susceptibility, T :
2 N
X
(2:10)
3
!2
T = N 4h N1 i i ; hmi2 5
i=1
the specic heat, CV :
" 2
H
]
CV = N h N i ; hei2
(2:11)
#
(2:12)
and many others.
In general, these quantities dier considerably from the mean eld values. In
Figure 4 we can see the dierence between the magnetization computed from
a numerical calculation in two dimensions and the mean eld result (which is
independent of dimension).
It is known that the 4 model belongs to the universality class of the Ising
model (Amit 1984). This means that both models share the same critical exponents and scaling functions. Here, however, we want to point out a dierent
relation between the two models, namely, that the Ising model can be obtained
as a suitable limit of the 4 model. The limit is obtained as follows: expanding
the square in the "gradient" term and using the periodic boundary conditions,
one can rewrite the Hamiltonian in the form:
N X
X
u
b
2
4
(2:13)
H(]) =
dK ; 2 i + 4 i ; K ij
i=1
hij i
where the sum over hi j i means sum over all the possible pairs of i{j nearest
neighbours in the lattice. Introducing a rescaled eld variable
Si = q i
(2:14)
b;2dK
u
one can write:
H(S]) = N
X
i=1
; 12 Si2 + 14 Si4 ; J
X
<ij>
Si Sj
(2:15)
Computational Field Theory and Pattern Formation
21
Fig. 4. Mean eld magnetization, Eq.(2.5) (solid line) and the result of a numerical
calculation using Monte Carlo techniques in the two{dimensional system (dashed line).
In this plot we use the value = 1 (see Eq.(2.21) for a denition of the parameters).
with
J = K(b ;u2dK)
= (;b +u2dK) The Gibbs factor is
2
8
N <
Y
(2:16)
39
2
=
X
1
1
2
4
;
H
(
S
])
Si Sj 5
e
= :exp ;(; 2 Si + 4 Si ) exp 4J
<ij>
i=1
(2:17)
The local factor of this distribution is always double peaked centered around
the values 1 (see Fig.5). Now we can think of the parameter as controlling
the width of the local eld distribution around these two maxima. A large value
for implies that the eld is bound to take essentially the values 1, and Si
becomes what is called a spin variable. In the limit ! 1 with J kept constant
we recover the Ising model. This limit is obtained in the original 4 model by
taking
b!1
u!1
(2:18)
b ! constant
b
J = K u
u
In this limit, one obtains for the Gibbs factor:
e;H(S ]) =
N
Y
i=1
This is equivalent to writing:
2
(Si2 ; 1) exp 4J
X
<ij>
3
Si Sj 5
(2:19)
22
Raul Toral
H(S]) = ;J
X
<ij>
Si Sj
Si = 1
(2:20)
Which is nothing but the Hamiltonian of the Ising model. For smaller values of
, the eld can uctuate around the two maxima, hence the name of soft spins
model that sometimes is given to the 4 model in this situation (Cowley 1980
Bruce 1980 Bruce et al. 1980).
Fig. 5. Function f (x) = exp ;(; 21 Si2 + 14 Si4 ) appearing in Eq.(2.17), in the cases
= 1 (solid line), = 10 (dashed line), = 100 (dotted{dashed line). When ! 1,
the function f^x(x) tends to the sum of two delta functions located at x = 1.
2.2. Monte Carlo methods
Before we proceed further and show how one can obtain numerically the quantities of interest in the 4 model, we notice that one of the three parameters
of the model, u b K, is redundant since it can be absorbed in the eld scale.
Many ways of redening parameters have been used in the literature (see (Toral
et al. 1990) for a review). We will use here a simple reparametrization in which
the eld is rescaled by a factor K 1=2 hence yielding a parameter independent
interaction term (this rescaling is obviously valid only in the ferromagnetic case
K > 0). Specically, we introduce a new eld , and two new parameters and
by the denitions:
= K 1=2 = Kb
(2:21)
= Ku2
The Hamiltonian, in terms of these new variables, can be written as:
Computational Field Theory and Pattern Formation
H(]) =
=
N
X
i=1
N
X
i=1
"
"
d ;
; 2 + 4 + 1 X
i ; i 2
2 i 4 i 2
=1
2d
~ 2 + 4 ; X
ii
i
i
2
4
=1
#
23
#
(2:22)
(the sum over runs over the 2d nearest neighbours of site i). Here we have
introduced
~ = 2d ; (2:23)
This Hamiltonian can be explicitly separated into the local part and the interaction part, H = H0 + HI with
#
N " ~
X
2
4
H0 =
i + 4 i
(2:24)
i=1 2
HI = ;
X
<ij>
i j
(2:25)
Now it is about time we start applying what we learnt in the previous chapter
about Monte Carlo techniques! To implement the Metropolis algorithm we need
to take the following steps:
(i) Select (randomly or sequentially) one of the N eld variables, i .
(ii) Propose a change to another close value, 0i , chosen randomly in the interval
i ; i + , with a suitable value selected such that the acceptance
probability is around 0:5.
(iii) Compute the change in energy, H that this proposed change produces.
(iv) Accept the proposed value 0i with probability min(1 e;
H ).
In computing the change of energy one does not have to use the full expression
(2.22), but rather notice that most of the terms disappear when subtracting the
old and the new energy, such that the change in energy is simply:
2d
~
X
H = 2 (0 2i ; 2i ) + 4 (0 4i ; 4i ) + (0 i ; i ) i
(2:26)
=1
This simple Metropolis algorithm has some convergence problems due to the
fact that most of the trials belong to a small probability region. This is particularly true in the vicinities of the Ising limit in which the values around 0
have a very small probability (see again Fig.5). Several methods have been proposed as an alternative to the Metropolis algorithm and all of these methods
t into the proposal/acceptance scheme developed in the previous chapter: use
a proposal probability density function gx^ (xjy) (where x (01 : : : 0N ) and
y (1 : : : N ) stand for the complete eld conguration after and before the
proposal, respectively) and an acceptance probability h(xjy), satisfying detailed
balance:
;H(x)
;H(y)
(2:27)
gx^ (xjy)h(xjy) e Z = gx^(yjx)h(yjx) e Z
24
Raul Toral
As usual, only one variable will be updated at a time, what makes the proposal
gx^ (xjy) become a function gx^ (0ijy). In the approach of reference (Milchev et al.
1986) the proposal gx^ (0i jy) is chosen independent of y and proportional to the
Gibbs factor of the local term of the Hamiltonian:
~
gx^ (0i ) = C exp(; 2 02i ; 4 04i )
(2:28)
this is a one-variable probability density function that can be sampled by any
of the methods explained in chapter 1. The authors in reference (Milchev et al.
1986) chose a numerical inversion method. Once this proposal has been taken,
in the detailed balance condition only the interaction part of the Hamiltonian
appears explicitly,
h(xjy)e;HI y] = h(yjx)e;HI x]
(2:29)
For this equation we can use the Metropolis solution:
h(xjy) = min(1 e;
HI )
(2:30)
where the novel feature is that only the change of the interaction energy appears
in the exponential. This adopts a very simple form:
HI = (0 i ; i )
2d
X
=1
i
(2:31)
In general, this procedure can be used when the Hamiltonian can be splitted in a
sum of local terms plus an interaction term. One can chose the new value of the
variable, independently of the old value, according to the distribution dictated
by the local term. This proposal is then accepted with the Metropolis probability
using only the interaction term.
The heat-bath algorithm has also been used for this model (Toral et al. 1990).
Let us remember than in the heat-bath algorithm, the acceptance probability is
equal to 1 (i.e. the proposal is always accepted), one variable 0i is changed at
a time, and the proposal probability density function gx^(0 i ) is obtained from
a distribution in which all the other variables remain constant, i.e. we need to
identify exactly where does the variable 0i appears, all the other terms will be
considered constants. From Eq.(2.22) it is very easy to nd out the expression
for gx^(0 i ):
!
2d
~ 02 04 0 X
0
gx^ ( i ) = A exp ; 2 i ; 4 i + i i
(2:32)
=1
A is some normalization factor depending on the values of the other elds (which
are considered to be constant in the heat{bath method). To sample this onevariable probability density function it will not be useful to use a numerical
inversion method, because this function depends of the sum of the 2d neighbours
of site i, which vary from trial to trial. To sample P
this distribution we can use
instead a rejection technique. If we introduce c = 2d=1 i , we can write the
above function as:
Computational Field Theory and Pattern Formation
i
h
1p exp ; (0i ; c=~)2
gx^(0 i ) = B exp ; 4 04i
2~;1
~;1=2 2
"
25
!#
(2:33)
B is another
normalization constant. This is the product of the function
exp(; 4 04i ) and a Gaussian distribution of mean c=~ and variance ~;1 (for this
we require ~ > 0). This Gaussian distribution can be generated by the relation
0i = c=~+ ~;1=2ran g() where ran g() is a Gaussian distributed random variable
of mean 0 and variance 1. Finally, the value 0 i is accepted with a probability
exp(; 4 04i ).
Another elegant way of implementinga Monte Carlo method for the 4 model
is that of Bruce (Bruce 1985) in his study of the border model (this is nothing but
the 4 model in the case ~ = 0). In his approach, the proposal gx^ (xjy) gx^(0 i )
is also independent of the old conguration y (in the same vein than heat-bath),
but gx^(0 i ) is chosen to be the sum of two Gaussians which best approximate
the actual local distribution of the eld 0i :
1 exp ; (0 i ; 1)2 + p1 exp ; (0 i ; 2 )2
p
2 1 2
212
222
2 2
gx^(0 i ) = 1
(2:34)
To sample this distribution one generates a Gaussian number, ran g() of mean
0 and variance 1 and a uniform number, , in (0 1). If < 0:5 then one chooses
the eld from the rst Gaussian distribution, i.e. 0i = 1 + 1ran g(), else, if
0:5, one chooses from the second, 0 i = 2 + 2ran g(). An initial guess of
the parameters 1, 2 , 1 and 2 is later rened with the information coming
from the computed local distribution. The acceptance probability must satisfy
detailed balance:
;H(y)
;H(x)
g (0 )h(xjy) e
= g ( )h(yjx) e
(2:35)
x
^
i
Z
x
^
i
Z
introducing the eective interaction Hamiltonian, H~ (]):
H~ (]) = H(]) +
N
X
i=1
ln gx^(i )
(2:36)
the detailed balance condition becomes:
h(xjy)e;H~(y) = h(yjx)e;H~ (x)
(2:37)
from which it is clear that a possible solution for the acceptance probability is
the Metropolis solution in terms of the function H~ :
h(xjy) = min 1 e;
H~
(2:38)
26
Raul Toral
2.3. Histogram extrapolation
We have now developed some tools to study numerically the 4 model. We shall
next apply them to the computation of some magnitudes of interest. Let us focus
rst on the magnetization m (see Fig.4). This is a function of the parameters and . If we are interested, say, in the variation of m with for a xed value of
, we have to run our numerical algorithm for several values of . For each run,
we have to make sure that the system has reached equilibrium (thermalization)
and then produce a sucient number of independent congurations such that
the errors are kept under reasonable limits. Repeating this procedure for a large
number of values in order to have smooth curves for the magnetization is a
painful and slow procedure. There is a way of improving somehow this process
by using information from simulations at a given value of the parameter to
obtain results for another value 0 . The idea, which is extremely simple (and
clever!), has been obtained and reobtained many times, but were Ferrenberg and
Swendsen (see Ferrenberg et al. 1989 and references therein) who dramatically
demonstrated its utility in the case of the Ising model. We now describe this
histogram extrapolation technique for the 4 model.
The congurations obtained in a numerical simulation using the Monte Carlo
method follow the Gibbs probability distribution which depend on the parameters and , that we now write out explicitly:
;H(])
f(] ) = e Z ( )
(2:39)
The type of averages one is interested on usually involve only the following
functions:
m0 =
m1 =
m2 =
m3 =
N
X
i=1
N
X
i=1
N
X
i
2i
(2:40)
4
i
i=1
N X
d
X
i=1 =1
(i ; i)2
other functions mi can be incorporated into what follows, but we restrict to
these for simplicity. Since the Hamiltonian depends only on linear combinations
of m1 , m2 and m3 , the probability distribution function in terms of m0 , m1 , m2
and m3 is:
;
1
2 m1 ; 4 m2 ; 2 m3 (2:41)
f(m0 m1 m2 m3 ) = N(m0 m1 m2 m3)Zexp
( )
Computational Field Theory and Pattern Formation
27
Here N(m0 m1 m2 m3)dm0 dm1 dm2dm3 is the number of congurations with
values of mi in the interval (mi mi + dmi ), i = 0 1 2 3. These values of
N(m0 m1 m2 m3 ) N(mi ]) are recorded during the simulation forming a
histogram (hence the name of histogram extrapolation). The partition function
can be expressed in terms of N(mi ]) as:
Z Y
1
(2:42)
Z ( ) =
dmi N(mi ]) exp 2 m1 ; 4 m2 ; 2 m3
i
The average value of any function G(mi]) can be computed as:
hG(mi ])i =
Z
Y
i
dmif(mi ] )G(mi])
(2:43)
The extrapolation scheme allows one to compute the probability density function
f(mi ] 0 0), for values 0 , 0 of the model parameters, if f(mi ] ) is known.
For our particular model, the method is based on the exact relation:
f(mi ] ) exp 2; m1 ; 4; m2
f(mi ] 0 0) = R Q
; m1 ; ; m2
dm
f(m
]
)
exp
i
i
i
2
4
0
0
0
0
(2:44)
Let us, in order to discuss the implication of this relation, consider the simpler
case 0 = for which the above formula reduces to:
f(mi ] ) exp 2; m1
0
R
; ;
(2:45)
f(mi ] ) = Q
i dmi f(mi ] ) exp 2 m1
This identity is telling us that if we know the probability distribution function
of the m0i s at a given value of the parameter then we know it for any other
value 0. This is wonderful! Only one simulation at a selected value of is
required! Once we have computed the probability density function f(mi ] )
we can obtain, by means of the previous formula, the probability density function
f(mi ] 0 ) and from the knowledge of f(mi ] ) we can compute any average
we need using Eq.(2.43).
In practice, alas!, things are not so easy. Relation (2.45) is indeed exact,
but what it is telling us is that the information contained in the tail of the
distribution of f(mi ] ) is used to compute the distribution f(mi ] 0 ), see
Fig. 6. The only problem is that the errors in the tail of the distribution usually
are very large, so making infeasible to extrapolate to a value of 0 very far from
.
Of course, the width of the distribution f(mi ] ) will be very important to
determine the extrapolation range . If the distribution is very wide, one can
extrapolate to farther away values than for a very narrow distribution. This is
the reason why this histogram extrapolation technique is useful near the critical
point where the distributions are wider.
0
0
28
Raul Toral
Fig. 6. Probability distribution functions f (mi ] ) at two dierent values of using
Eq.(2.43). According to Eq.(2.45) the region around the maximum of the function given
by the dashed line is obtained from the tail of the function given by the solid line, where
statistical errors are larger.
2.4. Finite size scaling
In Figures 7 to 10 we have plotted, for dierent system sizes and using the
histogram extrapolation scheme, results for several quantities of interest: magnetization, energy, specic heat and susceptibility. Each of these gures has been
produced by just two simulations whose location is marked with points in the
gures. The rest of the curves have been obtained using the extrapolation scheme
detailed above.
These gures show what is called nite size eects (Barber 1983). Let us
compare Fig. 7 for the magnetization with Fig.4 which depicts the mean eld
behavior and the expected true behavior. The theory predicts that there should
be a value of = c below which the spontaneous magnetization is strictly zero.
In fact, the magnetization behaviour near c can be characterized by a critical
exponent, such that:
(
0
if < c
m() =
(2:46)
a
if > c
Here = 1 ; c . This is a non-analytical behaviour which, strictly speaking, can
only appear in a system of innite size. For a nite system, the number of integrations appearing in the denition of the partition function and the calculation
of averages is always nite, thus leading necessarily to an analytical result. We
can understand intuitively why nite size eects will be more important near
a second order phase transition. In this situation the correlation length which
measures the linear range over which elds at dierent sites of the lattice are
correlated, diverges (in an innite system) with a power-law singularity:
Computational Field Theory and Pattern Formation
29
Fig. 7. Magnetization m for the two{dimensional 4 model for = 1 and dierent
values of the system side L = 8 16 24 32 40 (from bottom to top lines). For each
value of L a continuous line has been drawn by running two simulations at the points
marked with a dot and extrapolating to other values of using Eq.(2.45).
Fig. 8. Same than gure 7 for the energy e and = 0:1.
() jj;
(2:47)
For a nite system, the elds can not be correlated longer that the system side
and we must have L. The theory of nite size tells us exactly how (and
why!) the averages of interest behave (Cardy 1988). The basic idea is that now
the magnetization, say, becomes a homogeneous function of and the system
side L, m( L) = x m(=L).
~
The unknown exponent x is obtained by demanding
that in the innite system, and close enough to the critical point, one recovers
the known behaviour given by Eq.(2.46). This implies that the function m(z)
~
30
Raul Toral
takes a nite limit when z ! 0 and then:
x ~ ; x ;x
m() = Llim
!1 m( L) = m(0)
(2:48)
compared to (2.46) one concludes = ;x and then the prediction for the
magnetization near the critical point for a nite system is:
~ ; =c )L;1= ]
m( L) = ;= m(=L)
~
= L;= m(1
(2:49)
The typical way of checking this scaling behavior is to plot m( L)L= vs the
rescaled variable (1 ; =c )L;1= . If, as it is usually the case, the critical value
c and the critical exponents , are not known, this procedure implies a threeparameter t which is very dicult to do in practice. One can use the equivalent
scaling relations for the specic heat and the susceptibility:
~ ; =c )L;1= ]
C( L) = L= C(1
( L) = L= ~(1 ; =c )L;1= ]
(2:50)
Fig. 9. Specic heat for the two{dimensional 4 model for = 1. Same symbols and
meanings than in gure 7.
From Figures 9 and 10 we see that both C( L) and ( L) develop maxima at
locations c1 (L) and c2 (L), respectively. By using the above expressions for the
specic heat and the susceptibility one can see that these values behave as:
c1 (L) = c + a1 L;1=
(2:51)
c2 (L) = c + a2 L;1=
From where we deduce that a plot of c1 (L) and c2 (L) vs L;1= must yield a
straight line whose interception at the origin is precisely c . This can be seen
in Fig.11 for which the critical value c = 1:265 0:010 for = 1 is deduced.
Computational Field Theory and Pattern Formation
31
Fig. 10. Magnetic susceptibility for the two{dimensional 4 model for = 0:1. Same
symbols and meanings than in gure 7.
In this gure we have used the known value = 1 for the 2 ; d 4 model.
When the value of is not known one could use directly the three parameter t
ci (L) = c + ai L;1= to obtain c although the quality of the t usually is not
good enough to allow also for a very accurate measurement of . Once the value
of c is known one can use the remaining nite size scaling relations to obtain
the critical exponents.
3. Field Dynamics
3.1. Introduction
Up to now we have focused only on the equilibrium properties of the scalar
4 model. If one wants to write dynamical equations for the elds one nds
the diculty that the 4 Hamiltonian, so successful for equilibrium properties,
does not contain enough information (e.g. kinetic energy terms) to determine
the dynamics of the model. Instead of adding more terms to the Hamiltonian,
the usual approach to study dynamical properties is to consider that the eld
variables follow some phenomenological equations of motion. These equations
of motion will be stochastic in nature to reect our ignorance of the detailed
eect of the microscopic variables in the eld variables (Gunton et al. 1983b).
In addition, they must respect the basic symmetries of the system and lead, in
the limit of innite time, to the equilibrium canonical distribution given by the
Gibbs factor.
32
Raul Toral
Fig. 11. Pseudo critical parameters c1 (L) (open dots) and c2 (L) (black dots) dened as
the location of the maxima of the specic heat and susceptibility, respectively, plotted
vs. L;1 . The extrapolation to the origin, according to Eq.(2.51) yields the critical value
c = 1:265 0:010.
3.2. The Langevin equation
As discussed before, one writes down semi-phenomenological stochastic equations (Langevin equations) to model the time evolution of the eld. For the
simplest case of the 4 model without any further symmetry (remember than
the main symmetry of the model is the ! ;), one is guided by the fact that
at suciently low temperatures the nal conguration must be a minimum of
the Hamiltonian. Then, the time evolution is dened by a purely relaxational
model such that the speed of approaching the ground state is proportional to the
gradient of the Hamiltonian. One adds a stochastic contribution (usually called
the noise term) that ensures the correct thermal uctuations in the stationary
(equilibrium) state. The resulting Langevin equation is:
@(r t) = ;; H + (r t)
(3:1)
@t
(r t)
Here ; is a positive coecient usually assumed to be constant, although in some
cases one also studies the eect of the dependance of ; on the eld (Langer
et al. 1975 Lacasta et al. 1992, 1993). As in the equilibrium case, one assumes
a lattice discretization in which, instead of a real variable in every point of the
space, only points on a d-dimensional lattice are considered. In the lattice
case, the functional derivative H=(r t) becomes a partial derivative:
di (t) = ;; a;d @ H + (t)
(3:2)
i
0 @ (t)
dt
i
In the absence of the noise term, it is easy to show that the system evolves
towards a minimum of the Hamiltonian. We have:
Computational Field Theory and Pattern Formation
N @ H d
N
dH = X
i = ;;a;d X @ H
0
dt i=1 @i dt
i=1 @i
2
0
33
(3:3)
The statistical properties of the noise terms, i(t), have to be determined in
order to obtain the correct equilibrium distribution. They are supposed to be
Gaussian distributed of mean zero and correlations given by:
hi(t)j (t0 )i = Da;0 d ij (t ; t0 )
(3:4)
characteristic of the so called white noise (Gardiner 1985). D is a parameter, to
be determined, representing the intensity of the noise and, consequently, D must
increase with temperature. In the limit of the lattice spacing a0 going to 0, the
Kronecker-delta becomes a Dirac-delta function:
h(r t)(r0 t0)i = D(r ; r0 )(t ; t0 )
(3:5)
From now on, we will take the lattice spacing as unit of length a0 = 1. Since
i is Gaussian, these correlation functions completely determine its statistical
properties. We now relate the noise intensity D to temperature T. For later
reference, we will be more general and consider the set of stochastic dierential
equations:
N
di(t) = F (t) + X
Gij i (t)
(3:6)
i
dt
j =1
With the same statistical properties for the noise term than before, Eq.(3.4).
Gij is a constant matrix. The complete statistical description of the elds ]
is given by the joint probability density function P (] t), which satises the
Fokker-Planck equation (Gardiner 1985). The description given by the Langevin
equations is equivalent to the one given by the Fokker-Planck equation:
N
X
2
3
N
X
@P (] t) =
@ 4;F P + D H @P 5
i
@t
@
2 j =1 ij @j
i=1 i
(3:7)
N @
@ H P + D @P
@P (] t) = X
; @
@t
@
2 @i
i
i=1 i
(3:8)
The matrix H is H = GG> . In the case we are concerned with, Eq.(3.2), H = I,
unity matrix, and Fi is related to the partial derivative of the Hamiltonian:
It is easy to verify that the stationary solution of this equation is:
2;
Pst(]) / exp ; D H(])
(3:9)
Since we want the statistical stationary properties of ] to be determined by
the Gibbs distribution, exp ; H(])] we are led to identify:
34
Raul Toral
2; = =) D = 2k T;
(3:10)
B
D
D is then proportional to temperature, as we had anticipated. By using the
expression for the lattice Hamiltonian equation (2.1) we arrive at the following
equation:
di(t) = ;k T ;b ; u3 + K r2 + (t)
(3:11)
B
i
i
i
L i
dt
r2L is the lattice laplacian operator dened by:
r2L i =
2d
X
(i ; i)
=1
(3:12)
here the sum over runs over the 2d nearest neighbours of site i. By redening
time t ! t;KkB T and by introducing the same eld rescaling and parameters
than in the equilibrium case, Eq. (2.21), we arrive at the somewhat simpler
equation:
di(t) = ; 3 + r2 + (t)
(3:13)
i
i
L i i
dt
where the new noise variables i satisfy:
hi(t)j (t0 )i = 2ij (t ; t0 )
(3:14)
The Langevin equation we have considered so far is the simplest dynamical model than can be written for a scalar eld and in the famous taxonomy
of Hohenberg and Halperin (Hohenberg et al. 1977) is simply called model A.
Langevin-type equations can describe more complex systems, such as model B
which is obtained formally by replacing the constant coecient ; in Eq.(3.1)
dening model A by ;; r2. The resulting equation of motion is:
@(r t) = r2 ; H(]) + (r t)
(3:15)
@t
(r t)
where (r t) are Gaussian distributed random variables of mean zero and correlations:
h(r t)(r0 t0)i = ;D(t ; t0)r2 (t ; t0 )
(3:16)
We let the physical interpretation of this model to the next chapter. Here we
will mention that model B is suitable to describe the relaxational evolution
of a conserved eld, i.e. one for which the spatial integral of the eld is time
independent:
Z
m
dr(r t) ;! dm
(3:17)
dt = 0
The lattice discretization of model B equation leads to:
di(t) = r2 ; @ H(]) + (t)
i
L
dt
@(r t)
(3:18)
Computational Field Theory and Pattern Formation
35
Again, in order to ensure that the correct equilibrium distribution is obtained,
one needs to tune the correct value for D. The Fokker-Planck equation is now:
2
3
N @
N @ ;
@P(] t) = X
@H P + D X
4;; r2
;r2L ij P 5
L
@t
@
@
2
@
i
i=1 i
j =1 j
N
X
@ r2 ; @ H P + D @P
= ; @
L @i
2 @i
i=1 i
(3:19)
The equilibrium distribution is given by exp ; 2D; H(]) and we are led to
identify D = 2;kB T as in model A.
3.3. Numerical Solution of Stochastic Dierential Equations
The Langevin equations for models A and B that we considered in the previous
section can not be solved exactly. A great deal of what we know about the
behavior of the solution of these equations comes from numerical simulations.
In this section we want to study how to handle numerically Langevin equations
of the general form:
N
dxi(t) = F (t) + X
Gij i(t)
i
dt
j =1
i = 1 : : : N
(3:20)
Fi are functions that depend on the set of variables x1 : : : xN . If Gij also depend on x1 : : : xN , one talks of multiplicative noise and the resulting stochastic
dierential equations will be considered in the Stratonovich sense (van Kampen,
1981). If, on the other hand, Gij are constant functions, one talks of additive
noise. The noise variables i (t) are Gaussian distributed, independent random
variables of mean zero and correlations given by:
hi (t)j (t0 )i = Dij (t ; t0)
(3:21)
Of course, the appearance of the Dirac-delta functions tells us that we might
have some problems of mathematical rigor in we insist in considering the noise
variables i as simple functions. They should be considered as distributions, but
most of the results we can get with more mathematical rigor are also obtained
with the usual function formalism : : :and some care. In particular, since the noise
variables are not dierentiable, we can not use the standard Taylor expansions
that help so much in developing algorithms for numerically solving ordinary
(not stochastic) dierential equations (Press et al. 1986). Instead, we must use
integration expansions (Kloeden et al. 1992). We can express these ideas more
clearly with a simple example.
Let us consider a one variable stochastic dierential equation:
dx(t) = F (t) + (t)
(3:22)
dt
36
Raul Toral
where F(t) is a suciently smooth (dierentiable) function and (t) is a Gaussian
distributed variable of mean zero and correlations:
h(t)(t0 )i = D(t ; t0)
(3:23)
In fact, this equation is so simple that we can solve it exactly. However, let us
develop a numerical algorithm to solve it. The algorithm generates a recursion
relation that will allow us to compute x(t + h) given x(t), h is the integration
step. This is the same structure than in ordinary dierential equations in which
numerical methods do not give the solution for every value of time t but only at
regular intervals tn = t0 + nh, n = 0 1 : : : separated by a integration step h. In
order to nd the recursion relation we integrate (3.22) between tn and tn + h to
obtain:
Z tn+h
Z tn +h
Z tn +h
=
x(t
+
h)
;
x(t
)
=
dsF(s)
+
ds(s) (3:24)
dt dx(t)
n
n
dt
tn
tn
tn
the rst term of the right hand side of this equation can be approximated by
Z tn+h
dsF (s) = hF(tn) + o(h2 )
tn
(3:25)
The second term is a Gaussian variable since it is the integral of a Gaussian
variable. Hence, its statistical properties are completely determined by the mean
value and the correlations. Let us dene the random variable:
!h (t) =
whose mean value is:
h!h (t)i =
Z
and whose correlations are given by:
h!h (t)!h (t0 )i =
t+h
Z
t +h
Z
0
t
t+h
Z
ds(s)
(3:26)
dsh(s)i = 0
(3:27)
t
t+h
t
0
duh(s)(u)i =
Z
t+h
Z
t +h
0
duD(s ; u)
(3:28)
This integral is an easy exercise on delta-function integration. The result is:
t
ds
(
0
0
h!h (t)!h (t )i = D (h ; 0jt ; t j)
t
ds
t
0
jt ; t0j h
jt ; t0j h
(3:29)
The important thing to notice is that for the times tn that appear in the recursion relation (3.24), one has h!h (tn )!h (t0n )i = Dhnn , so that the variables
!h (tn ) are independent, Gaussian, random variables of zero mean and variance
h!h (tn )2 i = Dh. Hence they can be written as
0
p
!h (tn ) = Dhu(n)
(3:30)
Computational Field Theory and Pattern Formation
37
where u(n) are independent Gaussian variables of mean zero and variance unity
which can be generated, for instance, by the Box-Muller-Wiener method. The
nal recursion relation reads:
x(0) = x0
p
(3:31)
x(n+1) = x(n) + hF (n) + Dhu(n)
with the obvious notation x(n) = x(t0 + nh), F (n) = F(x(t0 + nh)), etc. This
simple example has taught us that the stochastic contribution is of order h1=2
whereas the deterministic contribution is of order h. This singular expansion is
the main reason for the failure of algorithms based on a na"ve Taylor series on
integer powers of h. Let us now consider another example:
dx(t) = F(x) + G(x)(t)
(3:32)
dt
Integration of this equation leads to (we have taken t0 = 0 to simplify notation):
x(h) = x(0) +
h
Z
0
dsF (x(s)) +
h
Z
0
dsG(x(s))(s)
(3:33)
To proceed, we Taylor expand the functions F(x) and G(x):
dF
(x(s) ; x(0)) + o(x(s) ; x(0))2
F(x(s)) = F(x(0)) + dx x=x(0)
2
(3:34)
G(x(s)) = G(x(0)) + dG
dx x=x(0) (x(s) ; x(0)) + o(x(s) ; x(0))
By substituting these expansions in (3.33) we notice is that at lowest order, h1=2,
we have
Z s
x(s) ; x(0) = G(x(0)) du(u)
(3:35)
0
and hence, to have an algorithm correct to o(h) it suces to consider the following terms in the expansion:
x(h) =x(0) + F(x(0))h + G(x(0))!h (0)+
Z h
Z s
(3:36)
3=2 )
G(x(0)) dG
ds
du(s)(u)
+
o(h
dx x=x(0) 0
0
p
The double integral gives !h (0)2 =2. Replacing !h by Dhu, see Eq.(3.30), we
arrive at the following recursion relation, known as the Milshtein algorithm (Milshtein 1974):
x(0) = x0
(n)
(3:37)
p
(n)]2
x(n+1) = x(n) + hF (n) + Dhu(n) + 12 G(n) dG
Dhu
dx
38
Raul Toral
The Milshtein method is an algorithm correct to order h and can then be considered as the stochastic equivalent of the Euler method to solve numerically
ordinary dierential equations. Sometimes in the literature (Greiner et al. 1988)
the name Euler method for stochastic dierential equations (in the Stratonovich
sense) is given to the above expression in which the term u(n)]2 is replaced by
its mean value hu(n)]2i = 1:
x(n+1) = x(n) + h
"
F (n) +
#
D G(n) dG (n) + pDhu(n)
2
dx (3:38)
This is an algorithm with a worst convergence that the Milshtein algorithm.
However, if one is interested only on computing the moments hx(t)k i of the
variable x(t) one can prove that this Euler algorithm has the same accuracy
than the Milshtein algorithm. Given that both algorithms have approximately
the same computational complexity, it does not seem justied to use the poorer
Euler algorithm instead of the Milshtein algorithm.
It is possible but very tedious to develop higher order algorithms. The good
news is that very rarely one really needs them. This is so because, in general,
one has to solve numerically the Langevin equations and average the results for
dierent realizations of the noise (and maybe initial conditions). This generates
a source of statistical errors coming from the averages which are, in many occasions, greater than the systematic errors due to the order of convergence of the
numerical method. So it is usually better to spend the computer time in reducing
the statistical errors by increasing the number of samples in the average rather
than using a more complicated, higher order, algorithm. However, we would like
to mention briey the stochastic Runge-Kutta type methods because, although
they do not yield better results than the Milshtein method for the convergence
of the stochastic terms, they do treat better the deterministic terms, increasing
in many cases the numerical stability. The simplest stochastic Runge-Kutta type
method is the Heun method which, for Eq.(3.32) reads:
k = hF(x(0))
p
l = Dhu(0)G(x(0))
(3:39)
y = x(0) + k + l
p (0)
x(h) = x(0) + h2 F(x(0)) + F(y)] + Dhu
2 G(x(0)) + G(y)]
One can prove by Taylor expanding the functions F (x) and G(x) that the Heun
method reproduces, to order h, the stochastic part of the Milshtein method (Gard
1987). For the purely deterministic case, D = 0, the Heun algorithm reduces to
the standard 2nd order Runge{Kutta method.
So far, we have only considered one-variable equations. It is very complicated
to develop algorithms for the general N-variable case of Eq.(3.20). The interested
reader can nd in (Kloeden et al. 1992) an extensive account of these general
methods. One can always use the following extension of the Euler method:
Computational Field Theory and Pattern Formation
2
x(in+1) = x(in) + h 4Fi(n) + D2
N X
N
X
j =1 k=1
(n)
ik G(jkn) @G
@x j
3
5
p
+ Dh
N
X
j =1
39
G(ijn)u(jn)
(3:40)
where the i are independent Gaussian variables of mean 0 and variance 1:
hu(in)u(in ) i = nn ii
(3:41)
In the case of diagonal noise, (Kloeden et al. 1992), i.e. one in which the noise
terms do not couple variable at dierent lattice locations:
Gij (x1 : : : xN ) = Gi (xi)ij
(3:42)
one can generalize the Milshtein algorithm to:
u(n)
0
0
p
x(n+1) = x(n) + hF (n) + Dhu(n) +
0
0
1 G(n) dGi (n) Dhu(n)]2
(3:43)
i
i
i
i
i
2 i dxi In next chapters we will show some applications of the numerical integration of
Langevin eld equations. In Chapter 4 the Langevin equation will be combined
with Monte Carlo techniques to yield a very useful tool to study equilibrium
properties: the hybrid Monte Carlo algorithm. In Chapter 5 we will be concerned
with the dynamics of growth in two very dierent situations: phase separation
and growth of random surfaces.
4. Hybrid Monte Carlo
4.1. Introduction
The numerical integration of Langevin equations described in the previous chapter has been used extensively as an alternative to Monte Carlo methods to sample
the canonical distribution. The idea is to run the Langevin equation up to the
stationary state (a process similar to thermalization for Monte Carlo) and then
average the dierent eld congurations produced over time. An advantage of
this procedure as compared to Monte Carlo is that it is not necessary to compute
the change in energy every time a single variable is updated. This is important
when the computation of the Hamiltonian is extremely time consuming. Such
situation arises in lattice gauge theories that include dynamical fermions (Zinn
Justin 1989). The eective Hamiltonian is so complicated for those systems that
it is prohibitive having to compute H N times per Monte Carlo step.
An advantage of Monte Carlo methods over the Langevin integration is that
the only errors in Monte Carlo are of statistical origin due to the limited number of congurations in the sample (and also to the correlations amongst them)
and are known to behave as the inverse square root of the number of congurations. On top of the statistical errors, the numerical integration of the Langevin
equation produces systematic errors due to the niteness of the time step used
in the integration. These systematic errors are much more cumbersome, since
40
Raul Toral
their numerical importance is not known a priori. A recent idea aims to combine
the best of both worlds. In the so called Hybrid Monte Carlo algorithm (Duane
et al. 1987) the numerical integration of the stochastic dierential equation is
used as the proposal to which an acceptance/rejection step is applied. Before
we develop this beautiful idea, we need to develop some concepts related to the
microcanonical ensemble.
4.2. Molecular Dynamics
Molecular dynamics (Allen et al. 1987) oers a direct numerical approach to the
behavior of a system with many degrees of freedom. Let X] = (x1 x2 : : : xN )
be microscopic variables and P] (p1 p2 : : : pN ) their conjugate momenta.
The Hamiltonian function is:
2
H(X P) = V (X) + T (P ) = V (X) + P2 = V (X) +
N
X
p2i
i=1 2
(4:1)
The representative point of the system x p] evolves in phase space according to
Hamilton equations
dxi = p
i
dt
(4:2)
dpi = F = ; @ H
i
dt
@xi
and the movement is conned to the hypersurface of constant energy, H(X P) =
E. This property can be used to perform averages on the microcanonical ensemble according to the ergodic principle that allows substitution of ensemble
#
averages hGi by time averages G:
Z
1
G# = Tlim
!1 T ; t
T
0 t0
dtG (x(t) p(t))
(4:3)
To perform this time average we can integrate numerically Hamilton equations.
In this way we can obtain equilibrium properties as well as dynamical (transitory) properties. However, similarly to what happened when using the numerical
integration of Langevin equations to sample the canonical distribution, two kinds
of errors are present: (i) statistical errors due to the nite number of samplings
and (ii) systematic errors due to the time discretization used in the numerical
solution.
The simplest integration scheme is the Euler method in which the derivative
of a function b(t) is approximated by:
db(t) = b(t + t) ; b(t) + O(t)
dt
t
Using this approximation, Hamilton equations become:
(4:4)
Computational Field Theory and Pattern Formation
xi (t + t) ; xi(t) + O(t) =p (t) 9
=
i >
t
pi (t + t) ; pi (t) + O(t) =F (t) >
i t
From where one gets the following recursive relations:
)
xi(t + t) =xi(t) + tpi (t) + O(t)2
pi (t + t) =pi (t) + tFi(t) + O(t)2
41
(4:5)
(4:6)
The local integration error is of order o(t)2 . After n integration steps, nt = t
the integration error is = O(n(t)2 ) = O(tt) = O(t). This error is usually
too large and demands choosing a very small time step for accuracy and numerical stability. Of course, there are many other higher-order integration methods.
Amongst all of them the leap frog algorithm will be particularly suited to our
needs. The algorithm is:
xi (t + t) = xi (t) + t pi (t) + t2 Fi(t)
(4:7)
t
pi (t + t) = pi (t) + 2 (Fi(t) + Fi (t + t)) i = 1 : : : N
and has a local error of order o(t)3 . Although this is an improvement over the
Euler method, still discretization errors show up, for instance, in the fact that the
energy is not exactly conserved. These energy uctuations (of numerical origin)
imply that the microcanonical ensemble is not sampled exactly.
Other properties of the Hamiltonian dynamics are, however, preserved by
this algorithm. In particular leap frog satises:
(i) Time reversal. If:
X(t) t X(t + t)
(4:8)
P (t) ! P(t + t)
then reversal of the momenta will take us back to the original value for X] and
the reversed value for the momenta P ]:
X(t + t) t X(t + 2t) = X(t)
(4:9)
;P(t + t) ! P (t + 2t)
;P(t)
(ii) Area preserving. The Jacobian of the change of variables induced during the
time evolution is equal to one:
X(t
+
t)
P(t
+
t)
=1
(4:10)
J
X(t) P(t)
But, let us repeat it once again, it does not conserve energy:
H = H(t + t) ; H(t) = O(t)m ] 6= 0
(4:11)
Summarizing, the numerical integration by the leap-frog method induces a mapping given by Eq.(4.7) that evolves the whole system from one point of phase
42
Raul Toral
space to another point with some error in the conservation of energy or, in
other words, leap-frog denes a global change of variables that almost conserves
energy. The idea naturally appears to combine this mapping with the acceptance/rejection technique typical of the Monte Carlo method. This is the main
idea behind the hybrid Monte Carlo method introduced by Duane et al. (1987)
that we now develop in more detail.
4.3. Hybrid Monte Carlo
Let us consider a scalar eld = (1 : : : N ) on the regular hypercubic lattice . The statistical properties of the system are given by the Gibbs factor
exp ;H(])] (we take = 1 to simplify notation), with a Hamiltonian H(]).
In order to dene a Hamiltonian dynamics we introduce some ctitious momenta elds Q] = (Q1 : : : QN ) and a new Hamiltonian function
2
H^ ( Q]) = H(]) + Q2 = H() +
Since
N
X
Q2i
i=1 2
(4:12)
i
2
exp ;H^ ( Q]) = exp ;H(])] exp ; Q2
(4:13)
from the statistical point of view, the momenta variables (Q1 : : : QN ) are simply
Gaussian distributed independent variables. By independent we imply not just
independent of each other but also of the elds ].
According to our plan, we want to make a proposal gx^(0 j) by using the
numerical solution of Hamilton equations. We specify some hinitial2 i values for
the momenta Q] according to the Gaussian distribution exp ; Q2 and then
integrate numerically, obtaining new values for the eld:
0
t (4:14)
Q ! Q0
This mapping can be done, for instance, by using the leap-frog algorithm,
Eqs.(4.7):
2
0i =i + t2 Fi + tQi
(4:15)
Q0i =Qi + t2 (Fi + Fi0)
We use this mapping to construct the proposal probability gx^ (0j). More specically, since the probability distribution function of the Q variables is exp(;Q2 =2)
we can write out the probability density function gx^ (0j) as:
h
gx^ (0j)d0 = dQe; 2
(4:16)
being Q the value of the momenta necessary to go from to 0. The key point
now is to choose an acceptance probability such that detailed balance is satised
with the hamiltonian H(]):
Q2
Computational Field Theory and Pattern Formation
43
e;H(]) gx^ (0j)h(0j) = e;H( ]) gx^(j0)h(j0)
(4:17)
The proposal gx^ (j0) is given again by integration of Hamilton equations with
a suitable value for the momenta, Q00]:
0
t (4:18)
Q00 ! Q000
from where one gets for gx^ (j0):
0
gx^(j0)d = dQ00e; 2
(4:19)
being Q00 the value of the momenta necessary to go from 0 to . The detailed
balance condition becomes:
Q00 2
2
Q 2
(4:20)
e; Q2 e;H() h(0j)ddQ = e; 2 e;H( ) h(j0)d0dQ00
Now the following properties apply:
(i) If the numerical integration satises time reversal, Q00 = ;Q0 hence the
Jacobian dQ00=dQ0 is equal to 1.
00
0
e; 2 e;H() h(0j)ddQ = e; 2 e;H( ) h(j0)d0dQ0
(ii) If it satises the area conserving property dQd = dQ0 d0
Q2
Q0 2
0
Q2
Q2
e; 2 e;H() h(0 j) = e; 2 e;H( ) h(j0)
Which, using Eq.(4.13), can be written as:
0
0
(4:21)
(4:22)
e;H^ (Q) h(0 j) = e;H^ ( Q ) h(j0)
(4:23)
We already know a possible solution to this equation, the Metropolis solution:
h(0j) = min(1 e;H^ )
(4:24)
^
^
^
where H = H(t + t) ; H(t)) is the change of total energy produced in
the evolution by time step t. Notice that, although we have made a global
change by updating all the variables i at once, H^ is controllable and the
acceptance probability can be made to stay within reasonable limits (close
to 50%). It is not necessary to make the acceptance/rejection decision every
time step t. Instead, usually one integrates during n time steps before
acceptance/rejection.
We give now an explicit implementation of the Hybrid Monte Carlo method:
(i) Generate independent Gaussian momenta Q].
(ii) Update the system Q] using leap{frog (or any other time reversible, area
preserving method) during a time t = nt
(t) t (t + t) t : : : t (t + t)
Q(t) ! Q(t + t) ! ! Q(t + t)
0
0
44
Raul Toral
(iii) Compute H^ = H^ (t + t) ; H^ (t)
(iv) Accept conguration 0 = (t + t) with probability min(1 e;
H^ )
Hybrid Monte Carlo constitutes an important tool for the simulation of eld
theories and it has been applied successfully to SU(n) eld theory (Sexton et
al. 1992), Lennard-Jones systems (Mehlig et al. 1992a), 4 scalar eld (Mehlig
et al. 1992b), X ; Y model (Gupta 1992), polymer systems (Forrest et al. 1994
Irbeck 1994), etc. and it is today a routine tool of great applicability.
There is another important aspect of the Hybrid Monte Carlo algorithm,
namely its relation with Langevin dynamics. Remember the Langevin equation
dening model A. Eq.(3.1):
@i = ; H + () = F () + ()
(4:25)
i
i
@
i i
where the noise terms satises:
< i()j ( 0) >= 2ij ( ; 0 )
(4:26)
The numerical solution using the Euler algorithm proceeds via the following
recursion relation:
p
(4:27)
i( + ) = i () + Fi() + 2ui ()
where ui () are Gaussian variables of mean 0 and correlations hui ()uj ()i = ij .
This is exactly the same evolution scheme of the leap-frog algorithm, Eq.(4.15),
if we make the following relation between the time steps of the two methods:
= t2 =2. We conclude that Hybrid Monte Carlo uses the same updating
scheme than Langevin (with a dierent unit of time), but the presence of the
acceptance/rejection step makes the method independent of the time step used
in the Langevin integration, so avoiding the systematic errors so dicult to deal
with in the Langevin method.
4.4. Generalized Hybrid Monte Carlo
It is clear that Hybrid Monte Carlo relies upon a numerical integration of the
dierential equations for a dynamical system in which the energy is conserved.
We will now introduce a general class of dynamical equations that also conserve
energy. This will give rise to a generalization of the hybrid Monte Carlo method
(Ferreira et al. 1993 Toral et al. 1994).
The generalization goes as follow: to each eld i we assign a vector momenta
variable Qi = (qi1 : : : qiD ) of arbitrary dimension D. We introduce the following
equations of motion:
D X
N
di = X
s
s
dt s=1 j =1(A )ij Qj
N
dqis = X
s
dt j =1(A )ji Fj
s = 1 : : : D
(4:28)
Computational Field Theory and Pattern Formation
45
or, written in more compact vector notation:
D
d = X
s s
dt s=1 A Q
(4:29)
dQs =(As )> F
s = 1 : : : D
dt
where the As are some linear operators which can be represented as a matrix,
and Fj represents the force as computed from the Hamiltonian ; @@ j H. We also
introduce a total Hamiltonian H^ , including the kinetic energy terms:
H^ ( Q]) = H() +
N X
D
X
(qis )2
i=1 s=1 2
(4:30)
The reader can easily verify that the proposed dynamics in Eqs.(4.28) exactly
conserves energy, i.e., dH^ =dt = 0. The standard hybrid Monte Carlo can be
obtained from the above dynamics setting D = 1 and A equal to the identity
operator.
For the approximate integration of the previous equations of motion the
leap{frog scheme can be used, introducing a discrete mapping (t) Q(t)] !
(t + t) Q(t + t)] = Gt ((t) Q(t)]), dependent on the time step t chosen.
The leap-frog approximation reads:
0 = + t
D
X
s=1
As Qs + (t)
2
D
2X
s=1
As (As )> F(])
Q0s =Qs + t (As )> (F(]) + F(0]))
(4:31)
2
It can be shown that this leap-frog approximation for arbritrary matrices As
satises the properties of time reversibility and area preserving. However, again
as a result of the time discretization used in the leap{frog scheme, the total
energy is no longer conserved although its variation can be controlled by varying t. We dene yet another mapping obtained iterating n times the previous mapping, i.e. G = (Gt )n. To satisfy detailed balance the conguration obtained when one applies G is accepted with probability min1 exp(;H^ )], where
H^ = H^ (G ( Q])) ; H^ ( Q]). As in the standard hybrid Monte Carlo, the
momenta variables are refreshed after every acceptance/rejection step according
to a Gaussian distribution of independent variables.
We have dened a general class of hybrid Monte Carlo-type methods characterized by a particular choice of matrices As . One can choose the matrices
As that better suit a particular problem. This generalized hybrid Monte Carlo
method turns out to be related to the method introduced in reference (Batrouni
et al. 1985) using the numerical integration of a Langevin equation with a matrix time{step. This method is based upon the observation that the stationary
solution of the Langevin equation (4.25) can be obtained approximately by the
recursion relation:
46
Raul Toral
i ( + ) = i() +
N
X
j =1
H + p2 p u
;ij ij j
j
(4:32)
Where ij is an arbitrary matrix and uj is a Gaussian variable of mean zero and
correlations hui uj i = ij . Comparing with (4.31) we can see that it corresponds
exactly to the one step leap-frog approximation of the generalized hybrid Monte
Carlo method introduced above taking D = 1 and if we identify: (t)2 =2 = and AA> = . The main dierence between the two methods is the presence
of an acceptance/rejection step in the generalized hybrid Monte Carlo. In this
sense, we can say that the generalized hybrid Monte Carlo method introduced in
this section makes exact (in the sense that averages are not biased by the choice
of the time step) the numerical integration of the Langevin equation using a
matrix time step introduced in reference (Batrouni et al. 1985).
4.5. Fourier Acceleration
Near a second order phase transition simulations become very time consuming
due to the large correlations which appear in the neighbourhood of the critical
point. This enhancement of the time correlations, known as critical slowing
down, greatly increases the statistical errors in a Monte Carlo simulationmaking
it necessary to average over a large number of congurations in order to obtain
values with an acceptable error. The physical origin of critical slowing down is
related to the fact that, near the critical point, uctuations of all wavelengths
are present, each with a characteristic time scale. In a Langevin integration,
for example, the integration step has to be small enough to treat correctly the
short{wavelength components but, on the other hand, the integration has to proceed for a sucient number of steps to have the long{wavelengths mode evolve
signicantly. In the Fourier acceleration method (Parisi 1983 Batrouni et
al. 1985), an integration technique is developed such that each mode evolves
with its own eective time step, large for long wavelengths and small for small
wavelengths. Whether Fourier acceleration is successful in completely overcoming critical slowing down for a particular system depends on to which extent the
evolution of the dierent Fourier modes can be considered indeed independent
of each other.
Fourier acceleration has been traditionally framed in the language of Langevin
equations. But we know already that numerical solutions of Langevin equations
introduce systematic errors due to the niteness of the time step used in the numerical integration. It is possible to develop exact Fourier acceleration methods
in the context of the generalized hybrid Monte Carlo algorithm developed in the
previous section. The basic ingredient is the following:
- Use matrices As whose elements are diagonal in Fourier space.
The optimal choice for these matrices depends on the system we are considering. Let us now discuss the example of the Gaussian model, for which the
Computational Field Theory and Pattern Formation
47
technique is most successful in reducing critical slowing down. Of course, the
example is purely academic, since the Gaussian model can be solved analytically
without having to resource to numerical methods.
The Gaussian model is dened by the following Hamiltonian for the scalar
elds i] (Parisi, 1988):
N
X
H=
2 + 1 j r j2
2 i 2 L i
(4:33)
i=1
index i runs over the N = Ld sites of a d-dimensional regular lattice , with
periodic boundary conditions rL is the usual lattice discretized version of the
gradient operator is a parameter called the eld mass. Although this model
does not show a real phase transition, the critical slowing down occurs for the
critical value" = 0.
To implement the generalized hybrid Monte Carlo method we choose the
number of momenta variables associated to a given eld equal to 1 (i.e. D = 1
in Eq.(4.28)). The total Hamiltonian H^ , Eq.(4.30), including the kinetic energy
terms can be written in terms of the Fourier transform of elds and momenta:
H^ =
N
X
k=1
!k2 j^ j2 + 1 jQ^ j2
2 k 2 k
(4:34)
the sum runsPover the N points in the reciprocal lattice, k = (k1 : : : kd ). Here
!k2 = + 4 dj=1 sin2(kj =2) and ^k and Q^ k stand for the elds and momenta
variables in Fourier space. Notice that the total Hamiltonian is a sum of terms
each one depending on a single mode k, such that the modes evolve independently
of each other.
According to the discussion above, we implement Fourier acceleration by
choosing the matrix A, generating the dynamics, diagonal in Fourier space. Let
us denote by A^k the diagonal elements of the matrix A in Fourier space. After
n leap-frog steps, the evolution equations (4.31) imply:
"
#
"
#
!k ^k (nt)
!k ^k (0)
n
= Mk
k = 1 : : : N
(4:35)
Q^ k (nt)
Q^ k (0)
The dynamical properties of the algorithm are determined by matrices Mkn, given
by:
sin(nk )= cos(k =2)
k)
Mkn = ; cos(cos(n
(4:36)
=2) sin(n )
cos(n )
k
k
k
where we have introduced k = cos;1 (1 ; c2k =2) and ck = A^k !k t. We see
explicitly how in this model dierent modes evolve independently of each other.
The fact that the evolution equations are linear greatly simplies the analysis of
this problem (this is similar to the standard hybrid Monte Carlo where A = 1,
Kennedy et al. 1991).
48
Raul Toral
We can now answer the question of which is the optimal choice for matrix
A. If we choose the following values for the diagonal elements of A in Fourier
space:
2
3;1=2
d
X
A^k = 1=!k = 4 + 4 sin2 (kj =2)5
j =1
(4:37)
the evolution matrices Mkn become independent of and k such all the modes
evolve with the same eective time step Ak !k t = t. This choice for A^k reduces
completely critical slowing down because correlation times (which are related to
the eigenvalues of Mkn) become independent of the mass even when tends
to zero and the model becomes critical.
5. Applications in Domain Growth
5.1. Dynamics of rst order phase transitions
5.1.1. The Cahn-Hilliard-Cook Equation
Many binary (A-B) mixtures which are homogeneous at high temperature T0,
phase separate when quenched below a certain critical value (Tc ) of the temperature. In the nal equilibrium conguration two phases coexist: the A and
B phases, each one rich in the A and B material respectively (see Fig.12). We
assume that the equilibrium properties of such mixture are given by the Gibbs
distribution at the nal temperature Tf and we are concerned here with the way
the mixture reaches this thermal equilibrium, i.e. with the dynamical process
of phase separation (Gunton et al. 1983 Binder 1990). In many occasions the
relevant equations of motion are those of model B dened in a previous chapter,
equations that we now justify from a more physical point of view.
The relevant eld in this case is a scalar eld representing the local concentration dierence of the two components of the mixture (r t) = A (r t) ;
B (r t). This eld obviously follows a conservation law:
Z
1
m V dr(r t) ;! dm
(5:1)
dt = 0
V
(V is the system volume). This global conservation law can be expressed in local
terms by means of a continuity equation:
@(r t) + rJ (r t) = 0
(5:2)
@t
The conservation current J (r t) provides the driving force for eld diusion.
Equilibrium is reached when the chemical potentials of the two substances are
equal and uniform. Thus, we write the current J (r t) as the sum of the gradient
of the dierence of chemical potentials of each component = A ; B (this is
the well known Fick's law) plus a random, noise type term to take properly into
account thermal uctuations:
Computational Field Theory and Pattern Formation
49
Fig. 12. Schematic phase diagram of a binary mixture A ; B . The system is homoge-
neous at temperature T0 . After quenching at temperature Tf below Tc the system phase
separates. At suciently late times, phase A and B coexist. The evolution proceeds
dierently if the quench is in region I or II , separated by the spinodal line (dashed
line).
J (r t) = ;; r + (r t)
(5:3)
; is a positive coecient called mobility. The noise variables = (1 : : : d )
are Gaussian distributed, of mean zero and correlations:
hk (r t)k (r0 t0)i = 2kB Tkk (r ; r0 )(t ; t0 )
(5:4)
The chemical potential is obtained as the derivative of H with respect to the
eld:
H(])
(r t) = (
(5:5)
r t)
Combining Eqs.(5.2) to (5.5), we reobtain model B equation, which in the context
of phase separation is called the Cahn-Hilliard-Cook (CHC) equation (Cahn et
al. 1958 Cook 1970):
@(r t) = r2 ; H(]) + (r t)
(5:6)
@t
(r t)
where (r t) = r (r t). If we adopt expression (2.1) for the Hamiltonian we
arrive at the following lattice form for the CHC equation:
di(t) = ;k T r2 ;;b + u3 ; K r2 + (t)
(5:7)
B L
i
i
i
L i
dt
If we perform the same eld and time rescaling than in the section 3.2 for model
A, we obtain the simpler equation:
0
0
50
Raul Toral
di(t) = r2 ;;b + 3 ; r2 + (t)
i
i
L
i
L i
dt
with correlations:
(5:8)
hi (t)j (t0 )i = ;2(t ; t0)r2L ij
(5:9)
The Cahn{Hilliard{Cook equation is a very complicated stochastic, nonlinear,
partial dierential equation that has so far deed an exact analytical treatment. Very successful systematic approximations valid for early times have been
developed (Grant et al. 1985). For the late stages of the dynamics the most successful theory is that of Langer, Bar-on and Miller (Langer et al. 1975), based
on some approximative mode decoupling. This theory, however, is very dicult
to improve upon due to its non-perturbative nature. It seems at this stage that
it is necessary to resource to numerical studies to analyze the long{time behaviour of the equation. One can integrate the CHC equation using any of the
stochastic numerical methods developed in section 3.3. In Figure 13 we have
plotted congurations resulting of a numerical integration using the simple Euler method. In this gure we observe that the evolution depends on the location
of the quench point in phase diagram. For critical (m = 0) or near critical concentrations, (region I of the phase diagram in Fig.12), the system is unstable
against long wavelength, small amplitude uctuations. This generates initially
a highly interconnected pattern that coarsens with time. The system is said to
evolve by spinodal decomposition. If, on the other hand, the system is in region
II, between the so{called spinodal line and the coexistence curve, the system
is unstable against the formation of nuclei of the minority phase. These nuclei
evolve in time in such a way that large droplets grow at the expense of small
droplets. This is the evolution for nucleation and growth. Although this simple
picture is thought to be of general validity, it is not possible to sharply separate
the domains in the phase diagram where the two processes dominate (Binder
1990 Chakrabarti 1992).
5.1.2 The Kawasaki model
A more microscopic approach to the dynamics of phase separation is the one
given by the Kawasaki dynamical version of the Ising model (Kawasaki 1972).
In each site of a regular lattice we dene two occupation variables, CiA and
CiB , such that CiA takes the value 1 (0) if in site i there is (there is not) an
A-atom. CiB is dened in a similar way. If we consider the situation in which
every site of the lattice is occupied by either A-atoms or B-atoms, i.e. that there
are no vacancies, CiA and CiB verify CiA + CiB = 1. We dene a new variable Si
as:
Si = CiA ; CiB
(5:10)
from where it is easy to deduce:
CiA = 1 +2 Si
CiB = 1 ;2 Si
(5:11)
Computational Field Theory and Pattern Formation
51
Fig. 13. Time evolution of the Kawasaki model and the Cahn{Hilliard{Cook (CHC)
model. Positive (negative) values for the elds or the spins are indicated by white
(black) regions. Notice the similarity between the Kawasaki model (left column) and
the CHC (center column) both at critical concentration (m = 0). In these two cases, the
system evolves by spinodal decomposition. In the right column, we plot a solution of
the CHC model for an o{critical quench (in this case m = 0:2). Notice that evolution
proceeds via the formation of nuclei of the minority phase that coarsen and grow.
Si is an Ising type variable taking the value +1 (;1) if in site i there is a Aatom (B-atom). Let us denote by Jij (AA), Jij (BB), Jij (AB) the strength of
the interaction between AA, BB and AB atoms, respectively, at locations i and
j in the lattice. The Hamiltonian is:
52
Raul Toral
H=
X
ij
Jij (AA)CiA CjA + Jij (BB)CiB CjB + Jij (AB)CiA CjB
(5:12)
Writing the occupation variable CiA and CiB in terms of the spin variables Si
one gets, apart from a constant factor:
H=;
X
ij
Jij Si Sj
(5:13)
Where Jij is some linear combination of the interaction strengths Jij (AA),
Jij (BB) and Jij (AB). We take now the approximation that the forces between
atoms are of short range nature, such that the only relevant interactions are the
ones between nearest neighbour sites on the lattice and we further assume that
they are all equal to a constant value J. In this case, the Hamiltonian reduces
to:
X
H = ;J
Si Sj
(5:14)
<ij>
where the sum runs over the nearest neighbour pairs on the lattice. This is nothing but the Hamiltonian of the celebrated IsingPmodel. In our binary mixture
model, though, there is the restriction that N ;1 i Si = m must be a xed constant. One can prove (Pathria 1972) that the grand canonical partition function
of the binary mixture model is equal to the canonical partition function of the
standard Ising model. Many equilibrium properties of the binary mixture, such
as the phase diagram, can thus be obtained from the corresponding ones of the
Ising model.
Our model for a binary mixture is indeed too simplied and it is absent of
many realistic features present in real materials, such as defects in the lattice,
no perfect symmetry between A and B compounds leading to asymmetric phase
diagrams, vacancies, etc. One can not pretend, then, to reproduce all of the
features present in these substances. In fact our hamiltonian is so simple that it is
absent of any \natural" dynamics such as the one given by Hamilton equations.
Kawasaki introduced an stochastic dynamics for the binary mixture model in
the same spirit that Glauber dynamics for the Ising model (Glauber, 1963). The
thermal uctuations, mediated by the lattice phonons, induce random transitions
in which atoms exchange positions in the lattice, see Fig.14 (in the ordinary
Glauber dynamics, spins ip randomly their value, this is forbidden here due
to the conservation law). This exchange process aims to mimic, in a very crude
way, the diusion process occurring in real materials.
In order to ensure that the system will asymptotically reach the correct
equilibrium state given by the Gibbs distribution, Kawasaki assigns a probability
to the microscopic exchange process. The equation governing the evolution of
the system is a master equation for the probability density of congurations
P(S1 : : : SN ] t):
d P(S] t) = ;P(S] t) X P(S] ! S 0 ]) + X P(S 0 ] t)P(S 0] ! S]) (5:15)
dt
S ]
S ]
0
0
Computational Field Theory and Pattern Formation
53
Fig. 14. Kawasaki exchange model basic process
The transition rate for the process 1 ! 2 (see Fig. 14) is taken in Kawasaki
formulation as:
(5:16)
P(1 ! 2) = 21 1 ; tanh 2kHT
B
Where H is the change in energy required to go from state 1 to state 2. simply xes the unit of time. The previous expression is chosen by mathematical
simplicity,but any function of the form P (1 ! 2) = exp ; k
BHT ! k
BHT where
!(z) = !(z ;1 ) will satisfy detailed balance for the master equation and hence
will ensure the correct equilibrium state (see discussion in section 1.6).
The Kawasaki exchange model and the Cahn{Hilliard{Cook model exhibit
many common features. In fact, it is possible to obtain, using some approximations, the Cahn{Hilliard{Cook model by means of a coarse{graining of the
Kawasaki exchange model (Langer 1971 Gunton et al. 1983b). In Fig.13 we
show the evolution of the CHC and Kawasaki models for the case of a quench
of critical concentration. Observe the great similarity between the evolution of
the two models.
5.1.3 Dynamical Scaling
An important property of the late stages of the dynamics of phase separation
process is that of dynamical scaling (Binder et al. 1974 Lebowitz et al. 1982).
This scaling description is suggested by the fact that systems at dierent times
look similar provided a space rescaling is performed. More precisely: if we take
a snapshot of a conguration of the system at a given time t1 and make it larger
by a factor it turns out that the resulting gure is statistically indistinguishable of a snapshot of the same system at a later time t2 = 1=a t1 , a being a
dynamical scaling exponent (see Fig.15). This dynamical scaling description is
based on the physical assumption that, at the late stages of the evolution, only
one length, related to the typical domain size, is relevant. Mathematically, the
scaling description implies that the pair correlation function G(r t):
54
Raul Toral
G(r t) = h(r t)(r + r0 t)i
(5:17)
or, rather, its circular average:
R
r t)
R
G(r t) = dG(
(5:18)
d
( denotes the angular variables), which, in principle, is a two variable function,
depends only on a combination of them, namely:
G(r t) = g(r=R(t))
(5:19)
where the scaling function g(x) is a one variable function of the scaling variable
x = r=R(t). R(t) is a measure of the typical domain size and can be dened, for
instance, as the rst zero of the pair correlation function. The time behaviour
of R(t) is described by a power{law:
R(t) ta
(5:20)
This expresses again the idea that the system at two dierent times t1 t2 will
look similar if a space rescaling of magnitude for the system at time t1 is
performed such that R(t1 ) = R(t2) or, according to the previous expression,
= (t2 =t1)a . Another important feature of the dynamical scaling description is
that of the universality of the scaling function g(x) and the scaling exponent a.
The value a = 1=3 seems to hold for the CHC and Kawasaki models independently of dimension and the quench location. The function g(x) does depend on
dimension and quench location (Chakrabarti et al. 1991 1993), although it seems
that its variation is very small for quenches close to the critical concentration
(Toral et al. 1989 Fratzl et al. 1989).
Dynamical scaling is usually checked in terms of the experimentally accessible
structure function S(k t), dened as the Fourier transform of the pair correlation
function G(r t). The scaling relation for S(k t) is obtained from the one holding
for G(r t) as:
S(k t) = R(t)d F (kR(t))
(5:21)
F (x) is the scaling function, d is the system dimensionality. In Figure 16 we
plot the time evolution of the angularly averaged structure factor S(k t) from a
numerical solution of the CHC equation in 3 dimensions in the case of a critical
quench. We can see that the structure factor starts from a rather at shape and,
as time goes on, develops a maximum whose location km moves towards smaller
values of k and whose intensity increases. The presence of this maximum signals
the phase separation process and the location of the peak is related to the inverse
typical linear length of the domains. In Figure 16 we also plot S(k t)R(t);d
versus kR(t) to check the scaling relation (5.21). One concludes that scaling is
well satised at the times studied. Similar results hold for other dimensions and
models, as well as for experiments (Binder 1990).
Computational Field Theory and Pattern Formation
55
5.2. Growth of rough surfaces
Many surfaces in nature grow roughly. Examples include such diverse systems
as biological tumors, mountain ranges, sand piles, a uid owing through a
porous medium and many others (Lagally 1990 Thomas 1982). To study this
widespread phenomenon, scientists have developed microscopic models and eld
models (Family et al. 1991).
The simplest of the microscopic models is one due to Eden (Eden 1958). The
Eden model can be used to study the spread of epidemics. In this model one
considers a regular lattice in which a site is initially marked as \infected". At a
given time a randomly selected infected site infects one of its nearest neighbours,
again chosen at random. After some time, the infected area is rather compact
with a rough surface, see Fig.17. From a careful analysis of this gure one can
deduce that the surface has fractal properties and that the dependence of the
roughness with time and with the linear dimensions of the Eden cluster can be
described by a power-law with some characteristic exponents.
Fig. 17. A cluster grown by the Eden model rules (see text). The right gure is a detail
of the cluster boundary to show that it is a rough surface.
Another kind of models aim to mimic the evolution of a surface that grows
by external ux of particles (Meakin et al. 1986). In the simplest solid on solid
model particles are added to a regular lattice. The incident particles add to
the height hi in lattice site i. The presence of a surface tension is simulated by
rejecting the addition of particles that produce a dierence in the height between
two nearest neighbours in the lattice larger that some xed amount (Kim et al.
1989).
One of the most successful eld theory models to study surface growth is
that of Kardar, Parisi and Zhang (KPZ) (Kardar et al. 1986). In the KPZ model
56
Raul Toral
Fig. 15. Self similarity after rescaling of congurations for the two{dimensional CHC
model in the case of a critical quench m = 0. A snapshot of the conguration at
time t1 = 50 000 (upper row, left) is magnied by a factor 256=119 approx2:15. The
resulting gure is very similar (from the statistical point of view) to a snapshot of
the same system at time t2 = t1 2:153 approx500 000, showing that the dynamical
exponent in Eq.(5.20) is close to 1=3.
Fig. 16. Raw data (left gure) for the structure factor for the time evolution of the
three{dimensional CHC in the case of a critical quench (m = 0). In the right gure
we have rescaled the date according to Eq.(5.21) to show that dynamical scaling holds
reasonably well for this system at the times studied.
Computational Field Theory and Pattern Formation
57
one denes a eld h(r t) giving the surface height at location r on a (d ; 1){
dimensional substrate at time t. Several physical processes contribute to the time
evolution of this eld.
(i) Tendency to grow normal to the surface. A simple geometrical argument
shows that in this case the evolution is given by:
@h(r t) = p1 + (rh)2 (rh)2
@t
2
to the lowest order in the gradient of the surface eld. Here is the surface
mean growth velocity.
(ii) Particle diusion on the surface.
r2h
is a constant related to the surface tension.
(iii) Random variations of the incident ux. These are modelled by a stochastic,
noise type term:
(r t)
Adding the contribution of the three terms we get the celebrated KPZ equation:
@h(r t) = r2h(r t) + (rh)2 + (r t)
(5:22)
@t
2
The (r t) noise is assumed to be Gaussian distributed of mean zero and correlations:
h(r t)(r0 t0)i = 2D(r ; r0 )(t ; t0 )
(5:23)
Some of the parameters of this equation, , , D are redundant. One can
reparametrize the eld h ! (=2D)1=2h and the time t ! t to obtain a simpler
equation:
@h(r t) = r2h(r t) + (rh)2 + (r t)
(5:24)
@t
2
where
h(r t)(r0 t0)i = (r ; r0 )(t ; t0 )
(5:25)
Although many exact results are known of this stochastic non{linear partial
dierential equation, an explicit solution is still missing. Numerical studies have
played an important role in characterizing the solutions of the KPZ equation
(Moser et al. 1991). In the numerical studies in a one{dimensional substrate one
typically uses the following discretization of the KPZ equation:
dhj (t) = hj +1 + hj ;1 ; 2hj + hj +1 ; hj ;1 2 + (t)
(5:26)
j
dt
a20
2
2a0
(periodic boundary conditions are coinsidered). The lattice spacing is usually
taken equal to 1, a0 = 1. The numerical integration of the previous set of stochastic dierential equations can proceed by any of the methods explained in Chapter
3, although the Heun method is particularly well suited in this case since it allows to use much larger values for the integration step t than the one allowed
58
Raul Toral
Fig. 18. Plot of the height eld h(x t) from a numerical simulation of the KPZ equation
in one{dimensional substrate for increasing times. We can see in this gure that the
surface grows rough and that the roughness increases with time. The right gure shows
a detail of the surface for the latest time.
in the Euler method (Forrest et al. 1993). If we write the previous equations in
the general vector form:
@ h(t) = F (h(t)) + (t)
(5:27)
@t
The Heun method uses the following recurrence relation (see Eqs.(3.39)):
g1 = F (h(t))
p
g2 = F (h(t) + tg1 + tu)
(5:28)
p
t
h(t + t) = h(t) +
2 (g1 + g2) + tu
Where u = (u1 : : : uN ) are independent Gaussian variables of mean zero and
variance unity.
In Figure 18 we can see the evolution of the eld for a 1-d substrate of length
L. It is clear from this gure that the surface grows rough, similarly to the case
of the microscopic Eden model. A quantitative measure of the surface roughness
is given by the quantity w(t L)
q
w(t L) = hh2 ; h2i
(5:29)
Where the bar denotes a spatial average h = L;1 Li=1 hi and the angular
brackets denote an average over initial conditions and realizations of the noise.
P
Computational Field Theory and Pattern Formation
59
It is important to characterize the time evolution of w(t L). The self{similar
character of the surface implies that the late{time behaviour for a system of
innite size must be given by a power{law relation of the form:
w(t 1) t
(5:30)
This relation has been well veried in computer simulations and the exact value
for can be obtained theoretically for 1-d systems ( = 1=3, see Familly et al.
1991). The reader could think that the verication of this simple law should be
an easy task. Not at all! In the rest of these notes, we just want to point out
which are the main diculties in trying to verify directly the, innocent looking,
power{law given by Eq.(5.30).
The rst and most obvious diculty comes from the fact that solving numerically a set of stochastic dierential equations is always dicult. One has
to average over a large number of realizations to have small statistical errors
and also make sure that results do not depend on the time step chosen for the
numerical integration.
The second problem is that expression (5.30) is an asymptotic result, i.e.
strictly valid only in the limit t ! 1. However, earlier time regimes can also be
described by a relation of the same functional form with an eective exponent
eff 6= . For example, at very earlyptimes, the noise term dominates (remember
that it gives a contribution of order t) and the roughness grows as the square
root of time, w(t L) t1=2. For intermediate times, the linear term dominates. It
is not dicult to solve this linear case ( = 0, the so{called Edwards-Wilkinson
model (Edwards et al. 1982)) and show that the roughness grows as w(t) t1=4.
It is only at the very late stages that the eect of the non{linearity fully develops
and one obtains the true asymptotic regime w(t) t1=3. But it is very dicult
to be sure that one is in the asymptotic regime and that the exponent obtained
from a, say, log{log plot of w(t) versus t is giving indeed the true asymptotic
exponent instead of an asymptotic exponent corresponding to smaller times.
Another problem concerns nite size eects. The result (5.30) is valid only in
the limit of an innite system. If the system size L is nite (as it is bound to be
in a numerical simulation), the surface roughness can saturate to an asymptotic
value w(t ! 1 L) L ( = 3=2 in 1-d) and one might never enter the time
regime in which relation (5.30) holds. The dramatic eect of the nite size eects
can be shown in detail for the linear case, = 0. In this case, the solution can
be found as:
L;1
2
2 (t L) = 1 X 1 ; exp(;8 sin (k=L)t) :
w(0)
(5:31)
L k=1
4 sin2 (k=L)
The strong L dependence of this linear solution can be seen in Fig. 19]. Looking
at the gure, we can say that only for L > 104 there is a relative error w(0) < 1%
for t ' 6000. The moral of the story is that if one wants to avoid nite size
eects one has to take systems at least at large as the ones needed in the linear
case, as it is sustained by a perturbative solution of the KPZ equation (Forrest
at al. 1993).
60
Raul Toral
2
Fig. 19. Dependence with L of the square of the surface roughness, w(0)
(t L) for the
linear solution of the KPZ equation, Eq.(5.31).
In fact, the nite size eects show up in a very determined way. It is now well
established (Family et al. 1985 Jullien et al. 1985) that the roughness w(L t),
for a nite substrate length L, follows an asymptotic scaling description in terms
of both time t and L:
w(t L) = t F (tL;= )
(5:32)
( is the so{called roughness exponent). In principle, this relation can also be
used to compute the exponent . Again, the diculty is that this relation is also
valid for the linear case, = 0 with dierent values for the exponents, namely
w(t L) = t1=4F (tL;2) and one has to be careful enough to ensure that the
non-linearity has fully developed. It is not strange, then, that the most powerful
available computers have been used in this problem of computing precisely the
exponents of the KPZ equation.
Acknowledgments
Finantial support from the DGICYT, project number PB-92-0046 is acknowledged.
References
Abramowitz, M. and Stegun, I.A. (1972): Handbook of Mathematical Functions, Dover
Publications.
Ahrens, J.H., Dieter, U. (1972): Comm. of the ACM, 15 873 (1988): 31 1331.
Allen, M.P., Tildesley, D.J. (1987): Computer Simulation of Liquids, Clarendon Press.
Amit, D.J. (1984):Field Theory, the Renormalization Group and Critical Phenomena,
World Scientic.
Computational Field Theory and Pattern Formation
61
Barber, M. (1983):Phase Transitions and Critical Phenomena, vol. 8, edited by C.
Domb and J.Lebowitz, Academic Press.
Batrouni, G.G., Katz, G.R., Kroned, A.S., Lepage, G.P., Svetitsky, B., Wilson, K.G.
(1985): Phys. Rev. D 32 2736.
Binder, K., and Stauer, D. (1974): Phys. Rev. Lett. 33, 1006.
Binder, K. (1990):Materials Science and Technology, Vol. 5 : Phase Transformation in
Materials, ed. P. Haasen, VCH (Weinheim, Germany), p-405.
Binder, K. and Heermann, D.W. (1988): Monte Carlo simulations in Statistical Physics,
Springer, Berlin, Heidelberg.
Binder, K. (ed) (1992):The Monte Carlo Method in Condensed Matter Physics, Topics
App. Phys. Vol 71 (Springer, Berlin, Heidelberg).
Box, G.E.P. and Muller, M.E. (1958): Ann. Math. Statist. 29, 610.
Bruce, A.D. (1980): Adv. Phys. 29, 111.
Bruce, A.D. and Cowley, R. (1980): Adv. Phys. 29, 219.
Bruce, A.D. (1985): J.Phys. A 18, L873.
Cahn, J.W. and Hilliard, J. E., (1958): J. Chem. Phys. 28, 258.
Cardy, J.L., editor (1988):Finite Size Scaling, North-Holland.
Chakrabarti, A., (1992): Phys. Rev. B 45, 9620.
Chakrabarti, A., Toral, R. and Gunton, J.D., (1991): Phys. Rev. A 44, 12133 (1993):
Phys. Rev. E 47, 3025.
Cook, H. (1970): Acta. Metall. 18 297 (1970).
Cowley, R. (1980): Adv. Phys. 29, 1.
Duane, S., Kennedy, A.D., Pendleton, B.J., and Roweth, D. (1987): Phys. Lett. B 195
216.
Eden, M. (1958):Symposium on Information Theory in Biology, edited by H. P. Yockey
(Pergamon Press, New York), p.359.
Edwards, S.F. and Wilkinsons, D.R. (1982): Proc. R. Soc. Lon. A 381 17.
Faires, J. and Burden, R. (1993): Numerical Methods, PWS Publishing Company,
Boston.
Family, F. and Vicsek, T. (1985): J. Phys. A 18, L75.
Recent reviews on kinetic roughening in surface growth models can be found in: F. Family and T. Vicsek, eds., Dynamics of Fractal Surfaces, (World Scientic, Singapore,
1991) J. Krug and H. Spohn, in Solids Far From Equilibrium: Growth, Morphology
and Defects, edited by C. Godreche, (Cambridge Univ. Press, Cambridge, England,
1991) L. Sander, ibid T. Vicsek, Fractal growth phenomena, (World Scientic, Singapore, 1989).
Feller, W. (1971):An introduction to probability theory and its applications, vols. 1,2.
John Wiley & Sons.
Ferrenberg, A.M. and Swendsen, R.H. (1989): Comp. in Phys. Sept/Oct, 101.
Ferreira, A.L. and Toral, R. (1993): Phys. Rev. E 47 R3848.
Forrest, B.M. and Toral, R. (1993): J.Stat.Phys. 70, 703.
Forrest, B.M. and Toral, R. (1994): J.Stat.Phys. 77, 473.
Fratzl, P. and Lebowitz, J.L., (1989): Acta Metall. 37, 3245 Fratzl, P. (1991): J. Appl.
Cryst.24, 593.
Gardiner, C.W. (1985): Handbook of Stochastic Methods, Springer-Verlag, Berlin.
Gard, T.C. (1987): "Introduction to Stochastic Dierential Equations", Marcel Dekker
inc, vol 114 of "Monographs and Textbooks in Pure and Applied Mathematics".
Glauber, R.J. (1963): J. Math. Phys. 4, 294.
Grimmett, G.R. and Stirzaker, D.R. (1982):Probability and Random Processes, Oxford
U. Press.
62
Raul Toral
Grant, M., San Miguel, M., Vi~nals, J. and Gunton, J.D. (1985): Phys. Rev. B 31, 3027.
Greiner, A., Strittmatter, W. and Honerkamp, J. (1988): J. Stat. Phys. 51, 95.
Gunton, J.D., San Miguel, M. and Sahni, P.S. (1983): in Phase Transitions and Critical
Phenomena, Vol. 8, edited by C. Domb and J. L. Lebowitz (Academic, London).
Gunton, J.D. and Droz, M. (1983): Introduction ot the Theory of Metastable and
Unstable States, Springer-Verlag.
Gupta, S. (1992): Nuc. Phys. B 370 741.
Heermann 1986: Computer Simulation Methods, Springer{Verlag.
Hohenberg, P.C. and Halperin, B, (1977): Rev. Mod. Phys. 49, 435.
Irback, A. (1994): J. Chem.Phys. 101 1661.
James, F. (1990): Comp. Phis. Comm. 60, 329.
Jullien, R. and Botet, R. (1985): J. Phys. A 18, 2279.
Kawasaki, K. (1972): in Phase Transitions and Critical Phenomena, vol. 2, edited by
C. Domb and M.S.Green, Academic Press.
Kalos, M.H. and Whitlock, P.A. (1986): Monte Carlo Methods, John Wiley and Sons.
Kardar, M., Parisi, G. and Zhang, Y.C. (1986): Phys. Rev. Lett. 56, 889.
Kennedy A.D. and Pendleton B. (1991), Nucl.Phys. B (Proc. Suppl). 20, 118.
Kim, J.M. and Kosterlitz, J.M. (1989): Phys. Rev. Lett. 62, 2289.
Kloeden, P.E. and Platen, E. (1992):Numerical Solution of Stochastic Dierential Equations, Springer-Verlag.
Knuth, D.E. (1981): "The Art of Computer Programming, vol.2 Seminumerical Algorithms", Addison-Wesley Pub. Co.
Lacasta, A.M., Sancho, J.M., Hernandez{Machado, A. and Toral, R. (1992): Phys. Rev.
B 45, 5276 (1993): Phys. Rev. B 48, 6854.
Langer, J.S. (1971): Ann.Phys. 65, 53.
Langer, J.S., Bar-on, B. and Miller, H.D. (1975): Phys. Rev. A 11, 1417.
Landau, L.D. and Lifshitz, E.M. (1980):Statistical Physics, Pergamon Press.
Lagally M. (ed.) (1990):Kinetics of Ordering and Growth at Surfaces, (Plenum, New
York), and references therein.
Lebowitz, J.L., Marro, J. and Kalos, M.H. (1982): Act. Metall. 30, 290 and references
therein.
Meakin, P., Ramanlal, P., Sander, L.M., Ball, R.C. (1986): Phys. Rev. A34, 5091.
Mehlig, B., Ferreira, A.L. and Heermann, D.W. (1992a): Phys. Lett. B291 151.
Mehlig, B., Heermann, D.W. and Forrest, B.M. (1992b): Phys. Rev. B 45 679.
Metropolis, N., Rosenbluth, A.W., Rosenbluth, M.N., Teller, A.H., Teller, E. (1953):
J. Chem. Phys. 21, 917.
Milshtein, G.N. (1974): Theory. Prob. Appl. 19, 557 (1978): ibid 23, 396.
Milchev, A., Heermann, D. and Binder, K. (1986): J. Stat. Phys. 44, 749.
Moser, K., Kertesz, J., and Wolf, D.E. (1991): Physica A 178, 215.
Pathria, R.K. (1972):Statistical Mechanics, Pergamon Press.
Parisi, G. (1984): in Progress in Gauge Field Theory (Cargese lectures), ed. G't Hooft
et al., Plenum Press (reproduced in Parisi, G. (1992), Field theory, disorder and
simulations, World Scientic.
Parisi, G. (1988):Statistical Field Theory, Addison Wesley.
Press, W.H., Flannery, B.P., Teulolsky, S.A. and Vetterling, W. (1986): "Numerical
Recipes". Cambridge University Press.
Rubinstein, R.Y. (1981): Simulation and the Monte Carlo method, John Wiley & Sons.
Sexton, J.C., Weingarten, D.W. (1992): Nuc. Phys. B380 665.
Sokal, A.D. (1991): Nuclear Physics B (Proc. Suppl.) 20 55.
Thomas, T.R. (1982):Rough Surfaces (Longman, London).
Computational Field Theory and Pattern Formation
63
Toral, R., Chakrabarti, A. (1990): Phys. Rev. B 42, 2445.
Toral, R., Chakrabarti, A. and Gunton, J.D. (1989): Phys. Rev. B 39, 901 (1992):
Phys. Rev. A 45, R2147.
Toral, R., Chakrabarti, A. (1993): Comp. Phis. Comm. 74, 327.
Toral, R. and Ferreira, A.L. (1994) Proceeding fo the Conference Physics Computing,
'94, p. 265, R.Gruber, M. Tomasini, editors.
Valls, O.T. and Mazenko, G.F. (1986): Phys. Rev. B 34, 7941.
van Kampen (1981): Stochastic Processes in Physics and Chemistry, North{Holland.
Zinn-Justin, J. (1989): Quantum Field Theory and Critical phenomena, Oxford. U.
Press.
This book was processed by the author using the TEX macro package from SpringerVerlag.